<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[Morphic]]></title><description><![CDATA[Security, systems thinking, and the patterns that hold everything together.]]></description><link>https://morphic.zenone.org</link><image><url>https://substackcdn.com/image/fetch/$s_!Iift!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66bf5ecd-c67a-4f99-b6b2-335fe455ccf1_1280x1280.png</url><title>Morphic</title><link>https://morphic.zenone.org</link></image><generator>Substack</generator><lastBuildDate>Wed, 18 Nov 2026 09:02:58 GMT</lastBuildDate><atom:link href="https://morphic.zenone.org/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Steve Zenone]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[morphic@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[morphic@substack.com]]></itunes:email><itunes:name><![CDATA[Steve Zenone]]></itunes:name></itunes:owner><itunes:author><![CDATA[Steve Zenone]]></itunes:author><googleplay:owner><![CDATA[morphic@substack.com]]></googleplay:owner><googleplay:email><![CDATA[morphic@substack.com]]></googleplay:email><googleplay:author><![CDATA[Steve Zenone]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[The AI Lab Is Getting Faster at Building AI]]></title><description><![CDATA[METR found that Anthropic's measured capability growth has nearly doubled. How much of that speed comes from AI helping build the next model?]]></description><link>https://morphic.zenone.org/p/the-ai-lab-is-getting-faster-at-building</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-ai-lab-is-getting-faster-at-building</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 11 Nov 2026 16:30:59 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!UV7d!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>METR estimates that Anthropic&#8217;s model capabilities on a set of measurable tasks have been improving about 1.9 times as fast as before October 2025.</p></li><li><p>METR thinks AI-assisted research explains much of the change. That causal judgment is less certain than the measured trend, and much of its supporting evidence was redacted.</p></li><li><p>Faster research could feed into better research tools. Whether that loop keeps accelerating remains an open question.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!UV7d!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!UV7d!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 424w, https://substackcdn.com/image/fetch/$s_!UV7d!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 848w, https://substackcdn.com/image/fetch/$s_!UV7d!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!UV7d!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!UV7d!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg" width="1456" height="832" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:832,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:399168,&quot;alt&quot;:&quot;Illustration of four glass flasks beneath a network of connected points in a blue-lit server room.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/jpeg&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/224203561?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Illustration of four glass flasks beneath a network of connected points in a blue-lit server room." title="Illustration of four glass flasks beneath a network of connected points in a blue-lit server room." srcset="https://substackcdn.com/image/fetch/$s_!UV7d!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 424w, https://substackcdn.com/image/fetch/$s_!UV7d!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 848w, https://substackcdn.com/image/fetch/$s_!UV7d!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!UV7d!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fea32845f-7553-44b4-a527-9eb8e5d4d4db_2688x1536.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">When AI helps researchers test more ideas, it may also help them build the next generation of AI.</figcaption></figure></div><p>Anthropic appears to have made roughly two years of its earlier benchmark progress in one. The researchers who tried to explain why were allowed inside the company to examine internal evidence.</p><p>Then Anthropic withheld much of the evidence those researchers considered most informative from the public report.</p><p>On November 5, the independent research group METR published an <a href="https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/">assessment of AI research and development at Anthropic</a>. Its team spent eleven weeks examining internal model evaluations, research data and interviews. METR&#8217;s central finding was that Anthropic&#8217;s capabilities on well-specified tasks appear to be improving at around twice the earlier rate.</p><p>The possibility behind that number is bigger than a faster model release cycle. AI systems are now helping researchers do the work that produces the next generation of AI systems. Better models could make that research faster again.</p><p>There are several steps between observing faster progress and establishing that AI is causing a self-reinforcing acceleration. The report is unusually insightful because it gives us evidence for some of those steps and admits how much remains uncertain.</p><h2>What actually got faster?</h2><p>METR analyzed scores from Anthropic&#8217;s internal models using AECI, Anthropic&#8217;s version of the Epoch Capabilities Index. It combines results from several capability benchmarks into one measure. METR estimated that progress was <strong>1.9 times as fast</strong> as before, with a reported 95% confidence interval of <strong>1.5x to 2.8x</strong>. That interval covers uncertainty in the model scores.</p><p>The word <em>acceleration</em> needs a little care here. METR isn&#8217;t saying each new model is 1.9 times as capable as the last one. It&#8217;s comparing how quickly the benchmark index improved across two periods. If a lab previously needed two years to make a certain amount of measured progress, a doubling of that rate would mean making the same progress in one year.</p><p>The report puts the change around October 2025. Its estimate includes internal models that hadn&#8217;t been released publicly, which is one reason it differs from the numbers in Anthropic&#8217;s published model cards.</p><p>The estimate also depends on how you construct the index. METR tried different choices of models, dates and benchmarks, with many results close to 2x. Changing the assumptions about how many benchmark tasks are easy or hard produced estimates from <strong>1.1x to 2.9x</strong>. Including models from other leading AI labs gave about <strong>1.4x</strong>, which changed the comparison: the earlier trend came mostly from other labs, while the recent trend came mostly from Anthropic.</p><p>So there is meaningful evidence that Anthropic&#8217;s measured progress sped up. The precise multiplier is less clear. And the index measures a particular collection of tasks, not every kind of reasoning, research or real-world impact.</p><p>I&#8217;ve written before about <a href="https://morphic.zenone.org/p/evaluating-agents-you-cant-trust">evaluating agents you can&#8217;t trust yet</a>. A benchmark can tell you something valuable about the tasks it measures without settling what happens when the work is messier. The same caution applies here. Faster progress on AECI is evidence of faster progress on AECI. How far that carries into open-ended scientific work is another question.</p><h2>A researcher can try more ideas</h2><p>Imagine a researcher who wants to try a change to a training method. Someone has to implement the experiment, connect it to the existing code, run it, inspect the results and decide whether to keep going. Some ideas take a day or two to set up. Others never get tried because the implementation work isn&#8217;t worth the likely payoff.</p><p>If an AI coding agent reduces that setup time, the researcher can test ideas that previously would have stayed on the whiteboard. The gain isn&#8217;t limited to doing the same work faster. It changes which experiments are affordable to attempt.</p><p>In interviews with METR, Anthropic researchers described trying substantially more ideas and shortening the time between an idea and an experiment. </p><p>METR also examined training compute, the computing resources used to train models, along with staffing, code contributions, experiment usage and other possible explanations. The team says it found convincing evidence that the acceleration wasn&#8217;t caused by faster growth in training compute. It couldn&#8217;t publish the underlying details.</p><p>After weighing what it saw, METR judged that AI probably accounts for a majority of Anthropic&#8217;s current acceleration. It offered a rough guess of about 80% for October 2026, with substantial uncertainty. The report also says non-AI factors likely played an important role in the acceleration that began in late 2025.</p><p>The benchmark trend is an estimate from observed scores. The share caused by AI is an informed judgment about what would have happened without AI. </p><h2>The 12x number is not the 1.9x number</h2><p>Anthropic staff responding to informal Slack polls estimated an average productivity gain of about <strong>12x</strong> from the most capable internal model in October 2026. METR treats the polls as very loose evidence because respondents may not represent the wider staff and the answers reflect little deliberation.</p><p>It&#8217;s tempting to put that beside 1.9x and declare that the numbers prove each other. They don&#8217;t measure the same thing.</p><p>The 12x figure comes from employees estimating how much faster they feel with AI. The 1.9x figure estimates the change in a benchmark growth rate. A researcher might get a coding task done much faster without the overall model-development process speeding up by the same amount. Training runs still take time. Experiments can fail. Someone still has to decide which result is worth pursuing.</p><p>In a <a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">2025 randomized study</a>, experienced open-source developers believed AI had made them faster. On the tasks studied, allowing AI actually made completion take <strong>19% longer</strong>. That result applied to early-2025 tools and those developers&#8217; particular work. METR now says it believes those historical results no longer reflect the current impact of AI models on open-source developer productivity.</p><p>METR tried a follow-up with newer tools, then <a href="https://metr.org/blog/2026-02-24-uplift-update/">reported in February 2026</a> that its design was struggling: developers who strongly preferred AI were less willing to participate in work without it, and some avoided submitting tasks they didn&#8217;t want to do manually. The researchers judged the newer results too biased to give a reliable overall speedup estimate.</p><p>I like that these findings sit beside the Anthropic assessment. The same research organization has evidence of a measured slowdown in one setting and probable acceleration in another. That&#8217;s not necessarily a contradiction. The people, tools, tasks and measurements are different.</p><h2>Independent judgment, evidence we can&#8217;t inspect</h2><p>METR wasn&#8217;t simply repeating an Anthropic press release. Its researchers received extensive internal access, examined multiple possible causes and reported disagreements and reversals in their own interpretations. METR says it wasn&#8217;t paid by Anthropic.</p><p>Anthropic had the right to redact non-public information before publication. METR says Anthropic removed <strong>a majority of the evidence it found most informative</strong> for its conclusions.</p><p>Anthropic said the redactions were &#8220;aimed at removing confidential information that could accelerate other AI developers.&#8221;</p><p>That includes important details about how research productivity was measured, which parts of the training process accelerated and how the team ruled out competing explanations. </p><p>I don&#8217;t think that makes the report worthless. Access to internal data can produce a better assessment than outsiders could make from public model releases alone. \</p><p>It does change what the word <em>independent</em> can reasonably promise. The evaluators made their own judgments, but the public can&#8217;t fully check those judgments against the underlying research.</p><h2>Does faster research keep making itself faster?</h2><p>Suppose an AI model helps researchers test twice as many ideas. Some of those experiments improve the next model. That model then helps researchers implement experiments that the previous one couldn&#8217;t handle.</p><p>There is a feedback loop in that description. Whether it keeps getting stronger depends on what else the research process needs.</p><p>Better coding assistance doesn&#8217;t by itself give the lab more computing power for training. It doesn&#8217;t guarantee that a promising result will work in other settings or that a lab can collect the data it needs. A researcher can generate more candidate ideas than the lab has capacity to test. At some point, the limiting step may move from writing the experiment to running it or making sense of the result.</p><p>METR looked for those constraints. It didn&#8217;t identify a clear bottleneck likely to stop further progress over the next twelve months, but the researchers say their assessment changed repeatedly as they received more information. They don&#8217;t claim to have proved that no bottleneck exists.</p><p>And METR says AI&#8217;s biggest contribution today is still <strong>assisting human researchers</strong>. Fully autonomous systems aren&#8217;t yet making major contributions to the most important parts of Anthropic&#8217;s training process, according to the information the team received.</p><p>That&#8217;s where I would draw the line on the more dramatic predictions. We have evidence that measured model progress has accelerated and a credible, though uncertain, case that AI-assisted R&amp;D is responsible for much of it. We don&#8217;t have public evidence that the whole research process can now improve itself without people, or that its rate of improvement will keep rising.</p><p>What I&#8217;d want to see next is fairly concrete: comparable capability measurements over more model generations, direct records of how much research work AI actually saves and independent access to the evidence connecting the two. I&#8217;d also want to know where extra experiments stop producing extra progress.</p><h2>Resources</h2><ul><li><p><a href="https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/">METR, &#8220;Independent Assessment of AI R&amp;D Acceleration at Anthropic in October 2026&#8221;</a>, November 5, 2026. Primary evidence, methods, redactions and limitations.</p></li><li><p><a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">METR, &#8220;Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity&#8221;</a>, July 10, 2025. Randomized developer-productivity study.</p></li><li><p><a href="https://metr.org/blog/2026-02-24-uplift-update/">METR, &#8220;We are Changing our Developer Productivity Experiment Design&#8221;</a>, February 24, 2026. Follow-up measurement limitations.</p></li></ul>]]></content:encoded></item><item><title><![CDATA[The Git Workflow Stayed. The Backend Had to Change.]]></title><description><![CDATA[OpenAI's Pando lets unrelated branches move independently and serves reusable Git history to more clients. Those choices come with tradeoffs.]]></description><link>https://morphic.zenone.org/p/the-git-workflow-stayed-the-backend</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-git-workflow-stayed-the-backend</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 04 Nov 2026 17:02:00 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!-u4j!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>OpenAI says its Git backend struggled as coding agents multiplied the number of branches, pushes and fetches hitting one monorepo.</p></li><li><p>Pando keeps the Git workflow but lets unrelated branches update independently. It also reuses stored Git data rather than building a fresh download for every client.</p></li><li><p>Pando doesn&#8217;t promise a single point-in-time view of all branches. That matters for tools that need a consistent repository-wide snapshot.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!-u4j!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!-u4j!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!-u4j!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!-u4j!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!-u4j!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!-u4j!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/a53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2210709,&quot;alt&quot;:&quot;Illustration of code symbols traveling along parallel tracks toward a cylindrical Git store.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/223492912?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Illustration of code symbols traveling along parallel tracks toward a cylindrical Git store." title="Illustration of code symbols traveling along parallel tracks toward a cylindrical Git store." srcset="https://substackcdn.com/image/fetch/$s_!-u4j!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!-u4j!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!-u4j!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!-u4j!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa53dddf3-d190-4e1e-a852-f42d30f9471a_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Same Git workflow, different machinery underneath. Pando lets independent branches move in parallel while reusing stored code history.</figcaption></figure></div><p>OpenAI says the service handling its monorepo, a single repository holding code for multiple projects, had a reliability problem. Fetches and pushes slowed. Some merges failed. At times, reliability fell below 90%.</p><p>Codex let an engineer work on several changes in parallel, each with its own branch, pushes and automated tests. That meant more systems asking the same repository for code, sometimes at the same time.</p><p>On November 2, OpenAI <a href="https://openai.com/index/scaling-git-at-openai/">published the engineering details of Pando</a>, the new backend now serving its monorepo. The company reports nearly 80% lower median push latency after switching, and says Pando delivers almost a petabyte of Git data daily. Those are OpenAI&#8217;s measurements for its own workload, not independent benchmarks.</p><p>Pando lets unrelated branches update independently, rather than requiring one update order across the whole repository.</p><h2>Why unrelated branches can end up waiting</h2><p>Picture one developer using three coding agents. One is updating a search feature. Another is fixing a payment form. The third is writing tests. Each agent creates commits on a separate branch, pushes them and asks continuous integration (CI) to test the result. CI then fetches the code, possibly from several machines.</p><p>The work on those branches might be completely unrelated. But a backend that orders every branch update through one shared coordination point can make them wait anyway. A faster network or a bigger machine won&#8217;t remove a queue that exists because every writer must take a turn.</p><p>Git is a protocol and a way of representing code history. How a hosted service stores that history, replicates it and handles competing writes is a separate engineering choice.</p><p>I&#8217;ve written about a related failure in <a href="https://morphic.zenone.org/p/coordination-is-an-architecture-problem">Coordination Is an Architecture Problem Disguised as a Prompt Problem</a>: two agents with conflicting views of who owns a task. That was about the shared state agents use to decide what to do. Pando is a narrower case. The agents can agree on their tasks and still run into a backend that makes independent code changes wait for one another.</p><h2>Pando changed two things</h2><p>First, Pando coordinates updates <em>per branch</em>, rather than requiring one ordered snapshot of every branch at once. A push to a search-feature branch can proceed while another client updates a payment-feature branch. If two clients try to move the same branch, they still have to resolve that conflict. The system hasn&#8217;t removed the need to coordinate. It&#8217;s reduced the scope of the coordination.</p><p>There is a technical name for part of this choice: <strong>causal consistency</strong>. When someone fetches a branch, Pando gives them the history that branch depends on. But if they fetch several branches while changes are arriving, those branches may reflect different moments in time. The system doesn&#8217;t promise one global snapshot for that read.</p><p>That sounds odd until you ask what the reader is actually doing. If a developer fetches one feature branch to review its commits, the history leading to that branch matters. Whether an unrelated branch changed a fraction of a second earlier may not. And if two features eventually need to work together, the merge and the tests still have to establish that they do.</p><p>Second, Pando avoids rebuilding a custom download every time someone fetches code. Git stores file contents and history as objects. Each object has an ID calculated from its type and contents. It can bundle those objects into compressed files called <em>packfiles</em>. Conventional servers often calculate which objects each client is missing, then generate a packfile for that particular fetch.</p><p>Pando saves incoming packfiles and reuses them. A client may receive some objects it already has, but the service can send the same stored bytes to many clients from caches instead of repeatedly doing expensive pack-generation work. Pando uses Git&#8217;s packfile-URI feature to send clients links to the stored packs they need. <a href="https://git-scm.com/docs/packfile-uri.html">The Git documentation describes this feature</a>, which requires protocol version 2. Pando still combines and recompresses packs in the background instead of doing that work separately for every fetch.</p><p>To make those packs reusable, Pando normally advertises only the trunk branch, the main development branch, as shared history when a client pushes a feature branch. Later pushes resend objects outside that trunk history, even if the server already has them. OpenAI reports that the median incoming packfile under this approach was less than 20 KiB and its 99th percentile was under 3 MiB. For its workload, the extra client bandwidth was worth accepting because another client with the trunk history could reuse the pack. That&#8217;s OpenAI&#8217;s observed tradeoff, not a universal rule.</p><h2>Testing and migrating the new backend</h2><p>OpenAI says Pando went from kickoff to production in ten weeks, with coding agents helping implement it. Engineers tested it against standard Git behavior, ran repeatable simulations to catch timing-dependent bugs and used formal models to check the design. Keeping familiar Git commands didn&#8217;t remove the need to test what happened when clients pushed and fetched at the same time.</p><p>The migration is almost as interesting as the design. OpenAI says it first ran Pando as a proxy in front of its old service, keeping both in sync while traffic moved over. The team then migrated hundreds of integrations that depended on earlier APIs, webhooks and downloads. A clever storage design doesn&#8217;t help much if the tools around it stop working during the switch.</p><p>The rollout included client changes too. OpenAI says it shipped its own Git build to developer machines and the ChatGPT desktop app, then used a wrapper around Git commands to move clients over in phases. The workflow stayed familiar, but the migration involved more than replacing the server.</p><p>On November 4, <a href="https://www.axios.com/2026/11/04/openai-pierre-computer-github">Axios reported</a> that OpenAI will acquire Pierre Computer, whose technology is already used in Codex. OpenAI told Axios it plans to integrate Pierre&#8217;s Code Storage more deeply. OpenAI separately says it&#8217;s exploring ways to share Pando&#8217;s technology with users, but its engineering post doesn&#8217;t announce a generally available service.</p><p>For teams watching these announcements, the question is less dramatic than whether Git is suddenly obsolete. OpenAI kept Git. It replaced the part underneath it that was making a particular workload expensive.</p><h2>What I&#8217;d measure before changing anything</h2><p>Teams with far less traffic than OpenAI may not need Pando. I&#8217;d start by looking at the delays already visible in the normal development process. When an agent finishes a change, how long does it wait for a push? How often do CI machines fetch the same large history? Are merges failing because of conflicting code, or because the service can&#8217;t keep up? Which operations still have to happen one at a time?</p><p>The answers matter because they point to different fixes. If most of the time is spent waiting for tests, faster pushes won&#8217;t change much. If the same Git history is repeatedly repacked and downloaded, caching or packfile reuse might help. If pushes to unrelated branches are waiting to update one shared record, the coordination design deserves engineering attention. If the real problem is conflicting edits to the same branch, letting unrelated branches move independently won&#8217;t solve it.</p><p>Before adopting Pando&#8217;s weaker cross-branch consistency, I&#8217;d identify every tool that needs a single snapshot across branches. A release system, for example, should record the exact commit IDs it will build and test. That makes the selected versions explicit, but it doesn&#8217;t prove they were all current at the same moment. If the release process requires that guarantee, it needs coordination elsewhere before giving up the repository-wide snapshot.</p><p>There&#8217;s a useful limit to what OpenAI&#8217;s numbers establish. They show a successful migration for one unusually large monorepo, using a backend tuned to OpenAI&#8217;s traffic. They don&#8217;t show that every Git service will slow down in the same way, or that every team should relax the same guarantees.</p><p>If more coding agents are making your work feel faster but your builds and code review still feel slower, follow the actual queue. In this case, the engineers found it underneath a familiar Git push.</p><h3>Resources</h3><ul><li><p><a href="https://openai.com/index/scaling-git-at-openai/">OpenAI Engineering, Scaling Git at OpenAI (November 2, 2026)</a>. Primary description of Pando&#8217;s architecture, deployment and company-reported performance.</p></li><li><p><a href="https://git-scm.com/docs/packfile-uri.html">Git documentation, packfile-uri</a>. Protocol reference for fetching packfiles by URL.</p></li><li><p><a href="https://www.axios.com/2026/11/04/openai-pierre-computer-github">Axios, OpenAI targets Microsoft&#8217;s GitHub with deal (November 4, 2026)</a>. Acquisition reporting, separate from the Pando engineering claims.</p></li></ul>]]></content:encoded></item><item><title><![CDATA[The Patient Couldn't Respond. The Brain Could.]]></title><description><![CDATA[Columbia researchers gave some unresponsive patients a way to change a sound using brain activity. It's not conversation yet, and a failed test proves less than you might think.]]></description><link>https://morphic.zenone.org/p/the-patient-couldnt-respond-the-brain</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-patient-couldnt-respond-the-brain</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 28 Oct 2026 15:31:07 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!TrSp!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>In an October 22 study, 13 of 41 clinically unresponsive patients used a noninvasive brain-computer interface to change a beeping sound.</p></li><li><p>The interface didn&#8217;t decode words or let patients hold a conversation. It detected a much simpler, deliberately controlled signal.</p></li><li><p>A negative result isn&#8217;t proof of unconsciousness. Even some patients who were clearly conscious couldn&#8217;t make the system work.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!TrSp!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!TrSp!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 424w, https://substackcdn.com/image/fetch/$s_!TrSp!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 848w, https://substackcdn.com/image/fetch/$s_!TrSp!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 1272w, https://substackcdn.com/image/fetch/$s_!TrSp!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!TrSp!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:277406,&quot;alt&quot;:&quot;Illustration of scalp electrodes connected to a signal display.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/222406361?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Illustration of scalp electrodes connected to a signal display." title="Illustration of scalp electrodes connected to a signal display." srcset="https://substackcdn.com/image/fetch/$s_!TrSp!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 424w, https://substackcdn.com/image/fetch/$s_!TrSp!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 848w, https://substackcdn.com/image/fetch/$s_!TrSp!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 1272w, https://substackcdn.com/image/fetch/$s_!TrSp!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd339bd6a-8bfc-4c17-9469-6d959261e6c9_1600x900.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">An EEG cap records electrical activity at the scalp while a computer looks for a deliberate response.</figcaption></figure></div><p>Imagine being asked to move your hand, understanding the instruction and being unable to make your hand move. If there&#8217;s no visible response, how do you know whether the person didn&#8217;t understand or simply couldn&#8217;t move?</p><p>That&#8217;s a question researchers at Columbia are trying to answer. In a <a href="https://pubmed.ncbi.nlm.nih.gov/42911846/">study published October 22 in </a><em><a href="https://pubmed.ncbi.nlm.nih.gov/42911846/">Brain</a></em>, they gave some patients with severe, recent brain injuries a different way to respond: change the timing of a beeping sound using electrical activity recorded from the scalp.</p><p>The patients were described as <em>clinically unresponsive</em>. That means their bedside examinations didn&#8217;t show the expected responses to commands. It doesn&#8217;t mean the researchers had established that every patient was unconscious.</p><p>Thirteen of the 41 patients managed to control the sound.</p><p>That&#8217;s more interesting to me than another headline about computers reading our minds. The researchers didn&#8217;t read anyone&#8217;s thoughts. They asked for a specific mental task and looked for a matching signal. In some patients, that gave them something the usual bedside examination couldn&#8217;t: evidence of a deliberate response.</p><h2>A bedside test can miss a person who&#8217;s trying</h2><p>When a doctor asks a patient to squeeze a hand or follow a moving finger, the response has to travel through several steps. The patient needs to hear and understand the request. They need to decide to act. Their brain must send the appropriate signals to the muscles, and those muscles have to move enough for someone to notice.</p><p>A patient with a major brain injury can fail somewhere along that path. If nothing moves, the doctor has observed a lack of movement. That observation alone can&#8217;t tell the doctor which step in the chain failed.</p><p>Researchers call one version of this problem <em>cognitive motor dissociation</em>. In ordinary terms, brain activity shows the patient following a command even though the bedside exam shows no response to that command. You may also see it called <em>hidden</em> or <em>covert consciousness</em>.</p><p>In a <a href="https://www.nejm.org/doi/10.1056/NEJMoa2400645">2024 study published in </a><em><a href="https://www.nejm.org/doi/10.1056/NEJMoa2400645">The New England Journal of Medicine</a></em>, researchers at six centers detected this kind of hidden command-following in 60 of 241 patients who showed no visible response to commands. They used brain scans or recordings of electrical activity. Unlike the new acute-care study, it included many people assessed months after their injuries. That&#8217;s 25% of the unresponsive group they studied, not 25% of every hospitalized patient.</p><p>Those tests could show that a brain responded to instructions. They didn&#8217;t give the patient a way to send a message back.</p><p>The new experiment asked a different question. Could a patient deliberately use the detected brain activity to control something, while the system was running?</p><h2>A beep, not a conversation</h2><p>Columbia&#8217;s team, led by neurologist Jan Claassen, used an electroencephalogram, usually shortened to <strong>EEG</strong>. Small electrodes on the scalp record patterns of electrical activity produced by groups of brain cells. There was no implanted chip and no surgery to connect the patients to the computer.</p><p>A patient heard a repeating tone through headphones. A recorded instruction asked the patient to try to open and close their right hand to make the beeps faster. To slow them down, the patient was told to stop trying that movement.</p><p>A computer analyzed the EEG signals as they arrived and adjusted the sound. The system also gave spoken feedback so a patient could hear when the change happened.</p><p>During the first part of training, the tone partly followed the expected response to the commands. The researchers counted control only in a later ten-minute test, when the decoder&#8217;s settings were fixed and it had to work from the EEG.</p><p><strong>The signal path:</strong> Spoken instruction &#8594; attempted hand movement &#8594; scalp EEG recording &#8594; computer analysis &#8594; changed beep.</p><p>The hand didn&#8217;t have to move. The system was looking for brain activity associated with attempting the task.</p><p>That&#8217;s what a <em>brain-computer interface</em> means here: software turns a measured brain signal into a limited action outside the body. In this case, it altered a sound. It didn&#8217;t translate a private thought, identify a feeling or figure out what the patient wanted for dinner.</p><p>The timing numbers need care, too. Across the ten-minute test, the researchers estimated how far the decoded EEG output lagged behind the command sequence. The <strong>median</strong> estimated delay was about <strong>50 seconds</strong> in the clinically unresponsive group and about 17 seconds in the conscious hospitalized group. That was an estimated lag across the test, not the measured time needed to answer a single question.</p><p>The researchers describe this as a <em>one-way</em> interface. Patients could change the tone, but the study didn&#8217;t demonstrate a reliable way for them to answer yes-or-no questions or tell a doctor where something hurt.</p><h2>The negative result is the part I&#8217;d be careful with</h2><p>The <a href="https://academic.oup.com/brain/advance-article-abstract/doi/10.1093/brain/awag210/8931697">published study</a> included three groups: 41 clinically unresponsive patients with acute brain injury, 22 conscious critically ill patients and eight healthy volunteers.</p><p>All eight healthy volunteers controlled the interface. But only nine of the 22 conscious hospitalized patients did. The other 13 couldn&#8217;t make the system work, despite being conscious.</p><p>That changes how I&#8217;d interpret a negative test. Failure could reflect fatigue, illness, fluctuating attention, trouble performing the task or difficulty detecting the signal. The study wasn&#8217;t designed to prove which explanation applied to every person.</p><p>Eight of the 41 unresponsive patients tested positive on a separate EEG-based test for hidden command-following. Thirteen controlled the tone. These weren&#8217;t interchangeable tests: <a href="https://academic.oup.com/brain/advance-article-abstract/doi/10.1093/brain/awag210/8931697">the study&#8217;s supplementary patient table</a> includes patients who succeeded on one but not the other.</p><p>That&#8217;s promising evidence that a different kind of task can reveal an ability another test missed. It isn&#8217;t proof that the interface can reliably tell whether any individual patient is conscious. A positive, deliberately controlled response is important evidence. An absent response leaves a lot unresolved.</p><p>I wouldn&#8217;t want a family to hear that their loved one &#8216;failed the consciousness test&#8217; when even some conscious patients couldn&#8217;t complete it.</p><h2>What the recovery numbers can and can&#8217;t tell us</h2><p>The researchers followed the patients after the experiment. By 12 months, six of the 13 unresponsive patients who controlled the tone reached the study&#8217;s definition of functional independence. That meant they could manage independently for at least eight hours in a day, based on a standard clinical outcome scale. Among the 28 who hadn&#8217;t controlled the tone, two reached that outcome.</p><p>The difference is striking, but the study followed patients at one medical center and relatively few people reached either outcome. The statistical uncertainty around the estimates was wide.</p><p>The study doesn&#8217;t show that using the interface improved recovery. The researchers observed an association between controlling the tone and later recovery. We don&#8217;t know whether using the device changed anyone&#8217;s outcome, or whether patients who were already more likely to recover were better able to use it.</p><p>The practical question is whether this kind of test could eventually give clinicians another piece of information when they&#8217;re assessing a severely injured patient. That&#8217;s a research direction, not a reason to make a treatment decision from this experiment alone.</p><h2>What would count as a real voice?</h2><p>The next challenge is turning control of the tone into answers a patient can choose and a doctor can reliably interpret.</p><p>For a patient to communicate, doctors would need a dependable way to establish what the patient intends, distinguish one answer from another and show that the answer came from the patient rather than a mistake in the system. They&#8217;d need to know how often it fails, how long it takes and what happens when attention or brain activity changes.</p><p>A tone that speeds up on command is evidence of controlled output. It isn&#8217;t yet a sentence, a preference or informed consent.</p><p>If a device someday offers a reliable communication channel, a patient may be able to express a need that nobody could previously hear. That would be extraordinary. But it would also make the accuracy of the channel matter enormously. An incorrectly decoded answer about pain is bad. An incorrectly decoded answer about a major medical choice is something else entirely.</p><p>The researchers themselves describe communication with families and clinicians as a future goal. Today&#8217;s experiment doesn&#8217;t provide it.</p><p>For now, I keep coming back to the bedside exam. A person doesn&#8217;t move. We see no answer. We may even be correct that the person cannot answer through the normal route. But a lack of movement doesn&#8217;t establish what that person&#8217;s mind can or can&#8217;t do.</p><p>The new experiment found a way for some of those patients to act without moving. For the others, we still don&#8217;t know enough.</p><div><hr></div><h3>Resources</h3><ul><li><p><a href="https://pubmed.ncbi.nlm.nih.gov/42911846/">Qi Shen and colleagues, </a><em><a href="https://pubmed.ncbi.nlm.nih.gov/42911846/">Clinically unresponsive patients with acute brain injury activate a brain-computer interface</a></em><a href="https://pubmed.ncbi.nlm.nih.gov/42911846/">, </a><em><a href="https://pubmed.ncbi.nlm.nih.gov/42911846/">Brain</a></em><a href="https://pubmed.ncbi.nlm.nih.gov/42911846/">, October 22, 2026</a>. Original peer-reviewed research abstract, methods, sample sizes, results and limitations.</p></li><li><p><a href="https://www.cuimc.columbia.edu/news/brain-computer-interface-may-give-unresponsive-icu-patients-voice">Columbia University Irving Medical Center, </a><em><a href="https://www.cuimc.columbia.edu/news/brain-computer-interface-may-give-unresponsive-icu-patients-voice">A Brain-Computer Interface May Give Unresponsive ICU Patients a Voice</a></em><a href="https://www.cuimc.columbia.edu/news/brain-computer-interface-may-give-unresponsive-icu-patients-voice">, October 22, 2026</a>. Researchers&#8217; description of the bedside experiment and its future goals.</p></li><li><p><a href="https://www.nejm.org/doi/10.1056/NEJMoa2400645">Bodien and colleagues, </a><em><a href="https://www.nejm.org/doi/10.1056/NEJMoa2400645">Cognitive Motor Dissociation in Disorders of Consciousness</a></em><a href="https://www.nejm.org/doi/10.1056/NEJMoa2400645">, </a><em><a href="https://www.nejm.org/doi/10.1056/NEJMoa2400645">The New England Journal of Medicine</a></em><a href="https://www.nejm.org/doi/10.1056/NEJMoa2400645">, August 2024</a>. Earlier multicenter study of hidden command-following.</p></li></ul><p><em>Source note: This article describes research findings, not a clinical test available for general bedside decision-making. A patient&#8217;s consciousness, prognosis and care require individual assessment by the treating medical team.</em></p>]]></content:encoded></item><item><title><![CDATA[The Video Showed a Yield Sign. The AI Recap Said Stop.]]></title><description><![CDATA[A new experiment found that one wrong detail in a recap changed what people later remembered seeing.]]></description><link>https://morphic.zenone.org/p/the-video-showed-a-yield-sign-the</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-video-showed-a-yield-sign-the</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 21 Oct 2026 15:30:49 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!BidR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>Researchers showed people a short animated accident, then deliberately gave some of them a recap that named the wrong traffic sign.</p></li><li><p>Among 328 people included in the analysis, 83.6% identified the sign correctly after an accurate recap, compared with 44.8% after a misleading one.</p></li><li><p>AI didn&#8217;t invent this effect. But if we rely on generated recaps, we need a better way to check them against what actually happened.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!BidR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!BidR!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!BidR!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!BidR!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!BidR!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!BidR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f8e927ff-2236-421c-9774-32f38a955da5_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2284740,&quot;alt&quot;:&quot;A yield sign at a neighborhood intersection, with a stop sign reflected in a mirror beside it.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/221224484?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="A yield sign at a neighborhood intersection, with a stop sign reflected in a mirror beside it." title="A yield sign at a neighborhood intersection, with a stop sign reflected in a mirror beside it." srcset="https://substackcdn.com/image/fetch/$s_!BidR!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!BidR!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!BidR!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!BidR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff8e927ff-2236-421c-9774-32f38a955da5_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Illustration, not a frame from the experiment.</figcaption></figure></div><p>Someone watches a short video of a car approaching a yield sign. The car turns and collides with a pedestrian. A day or two later, the person reads a recap of the video.</p><p>The researchers deliberately change the sign in the recap, so it says stop instead of yield. Then they ask the person which sign they remember seeing.</p><p>In a recent experiment, that kind of misleading recap made a big difference. People who read the incorrect version were much less likely to identify the original traffic sign correctly, even when the researchers told them to answer from their memory of the video rather than from the recap.</p><p>This is the part I find fascinating. We talk a lot about checking what AI writes. But what happens if reading the wrong description changes what the person doing the checking remembers?</p><h2>What the researchers tested</h2><p>A <a href="https://arxiv.org/html/2609.28820">paper posted September 23</a> by Mattea Sim, Yael Eiger and Tadayoshi Kohno looked at two related questions. First, how accurately do AI systems describe short videos? Second, can an inaccurate AI-generated recap affect what people remember about a video they already watched?</p><p>For the memory experiment, participants saw a short animated car-and-pedestrian accident. The key difference between the two video versions was the traffic sign at the intersection: stop or yield.</p><p>Between 24 and 48 hours later, participants returned and read a recap generated by ChatGPT. The researchers deliberately controlled one detail. Some people received a recap that named the correct traffic sign. Others received one naming the wrong sign.</p><p>That detail matters because the study wasn&#8217;t simply measuring whether a chatbot happened to make a mistake. The researchers intentionally changed the sign description to test its effect. All participants read AI-generated text, but some were told it came from AI and others were told it came from a human transcriber.</p><p>Of the 331 people who completed both parts, 328 passed the required checks and were included in the analysis.</p><p>The memory test asked participants to choose between stop and yield. They identified the original sign correctly at these rates:</p><ul><li><p><strong>83.6%</strong> of people who read the accurate recap identified the traffic sign correctly.</p></li><li><p><strong>44.8%</strong> of people who read the misleading recap identified the traffic sign correctly.</p></li></ul><p>The researchers also checked whether participants remembered other parts of the video. They generally did. For example, 96.6% correctly remembered that the accident happened during the day, a detail the recap hadn&#8217;t mentioned.</p><p>The mistaken answers weren&#8217;t simply proof that people had forgotten the entire video. The wrong sign in the recap had become part of how they answered a question about what they saw.</p><h2>This isn&#8217;t a new problem, but the delivery has changed</h2><p>Psychologists have studied something called the <em>misinformation effect</em> for decades. In plain English, information we hear or read after an event can affect how we remember the original event.</p><p>That&#8217;s why questions asked after an accident can matter. A person might be certain about what they saw and still have their memory nudged by later information.</p><p>The new study didn&#8217;t show that AI invented this weakness. The researchers didn&#8217;t find a statistically significant difference between people told the recap came from AI and people told it came from a human. An AI label wasn&#8217;t enough to prevent the effect in this experiment.</p><p>What has changed is how easy it is to produce and distribute a polished account of an event before anybody checks it carefully.</p><p>Meeting transcripts get shortened into action items. Calls become customer notes. Videos get turned into incident reports. A long conversation gets compressed into a paragraph someone else will read before opening the original.</p><p>Those are useful features. I&#8217;m not arguing we should stop summarizing things. I&#8217;m asking what happens when the convenient version contains a mistake and becomes the version people return to as the source of truth.</p><h2>Twenty summaries, and a reason to be careful</h2><p>The researchers also tested ChatGPT-5.5 and Gemini 2.5 Flash-Lite on two 25-second animated videos. They ran each model five times per video, using the same prompt each time. It asked for a factual, neutral account of at least 300 words.</p><p>All 20 summaries contained errors under the researchers&#8217; coding scheme. Nineteen omitted the collision with the pedestrian. Other mistakes included getting the order or direction of actions wrong.</p><p>Those results show how these recaps went wrong, but don&#8217;t establish a failure rate for other models, prompts, videos or meeting notes.</p><p>The memory experiment also compared accurate and misleading recaps, with no group that read no recap. The difference therefore doesn&#8217;t tell us how either version compares with relying on memory alone.</p><h2>A review can go wrong before anyone checks a fact</h2><p>Here&#8217;s a hypothetical work example.</p><p>An engineer joins an incident call. A colleague discusses a configuration change but doesn&#8217;t approve it. The next day, an AI-generated recap says the colleague approved the change at 2:15 PM.</p><p>The engineer was on the call. But after reading that recap, they may remember an approval that never happened. That&#8217;s the possibility I&#8217;d want to test in a real incident review, not a result this experiment established.</p><p>What concerns me is whether someone who already saw or heard the original can still catch the recap&#8217;s mistake from memory.</p><p>In <a href="https://morphic.zenone.org/p/the-source-is-becoming-optional">The Source Is Becoming Optional</a>, I wrote about how AI search can make us less likely to open the material behind an answer. This problem starts after somebody has already seen the source. A misleading recap may affect what they remember from it.</p><p>And in <a href="https://morphic.zenone.org/p/memory-is-a-write-path">Memory Is a Write Path</a>, I looked at what software stores and carries into later work. Here, the memory at risk belongs to a person.</p><h2>What I&#8217;d change about reviewing AI recaps</h2><p>When a recap could affect a decision, I&#8217;d change the order of a few ordinary steps. These are precautions I&#8217;d test, not methods this experiment proved work.</p><p><strong>Keep the original.</strong> A recording, transcript, log or document should remain available, with its source and date intact. A summary is a convenience. It isn&#8217;t a replacement for the thing being summarized.</p><p><strong>Give reviewers a chance to record what they saw first.</strong> Where practical, have a witness, investigator or meeting participant write a short account before they see somebody else&#8217;s polished recap. That preserves their first account for comparison.</p><p><strong>Check the claims that could change a decision.</strong> Who approved what? What time did it happen? Which person said something? Which action happened first? Go to the timestamp or original record for those details. Don&#8217;t ask a reviewer to sign off after simply reading the recap a second time.</p><p><strong>Make corrections visible.</strong> If the recap says stop sign and the video shows yield, record the discrepancy instead of quietly editing the final paragraph. Otherwise somebody may keep an older copy and repeat the error later.</p><p>There will be work where a fast summary is perfectly fine. Nobody needs a forensic review to summarize lunch plans. But an incident report, a legal account or a decision about someone else&#8217;s job deserves more care.</p><p>I like what AI summaries can do. But having seen the original earlier wouldn&#8217;t be enough for me to approve a recap that could affect a decision. I&#8217;d want the reviewer to compare it with the source, not just their memory of it.</p><h2>Resources</h2><ul><li><p><a href="https://arxiv.org/html/2609.28820">Sim, Eiger and Kohno, AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory, September 23, 2026</a></p></li><li><p><a href="https://osf.io/wmjtn">Study data and materials, Open Science Framework</a></p></li><li><p><a href="https://arxiv.org/abs/2609.28820">arXiv abstract and version history</a></p></li></ul>]]></content:encoded></item><item><title><![CDATA[Where Would You Shut an AI Worm Down?]]></title><description><![CDATA[Anthropic ran a model on cheap TV boxes. Its new research asks how close AI is to copying itself onto devices outside a provider's control.]]></description><link>https://morphic.zenone.org/p/where-would-you-shut-an-ai-worm-down</link><guid isPermaLink="false">https://morphic.zenone.org/p/where-would-you-shut-an-ai-worm-down</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 14 Oct 2026 17:03:46 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!YGjl!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>Anthropic tested AI running on cheap TV boxes and replication in a separate contained network. It didn&#8217;t demonstrate a complete AI worm spreading through homes.</p></li><li><p>If malicious software can run its own model, closing a model-provider account won&#8217;t stop that local copy. Defenders would need to isolate and clean up the machines running it.</p></li><li><p>Check which devices you own, whether they still receive updates and what they can reach on your network.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!YGjl!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!YGjl!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!YGjl!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!YGjl!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!YGjl!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!YGjl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/dc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:3785877,&quot;alt&quot;:&quot;Living-room illustration with a television, router and small devices connected by red lines.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/220394658?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Living-room illustration with a television, router and small devices connected by red lines." title="Living-room illustration with a television, router and small devices connected by red lines." srcset="https://substackcdn.com/image/fetch/$s_!YGjl!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!YGjl!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!YGjl!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!YGjl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc044e1a-d613-4ade-8433-b4a4813bcf25_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption"><em>Ordinary devices are still computers. The owner isn&#8217;t always the only person using them.</em></figcaption></figure></div><p>The unsettling part of an AI worm is where you&#8217;d have to go to turn it off.</p><p>A computer worm is software that copies itself from one machine to another. An AI agent uses a model to choose actions and tools to carry them out. Put the two together, and the software could bring a model along when it spreads. In a <a href="https://arxiv.org/html/2606.03811v1">June research preprint</a>, a paper released before peer review, Jonas Guan and colleagues tested an adaptive worm in a contained network. The software used a downloadable model rather than a commercial AI service.</p><p>If an agent needs a provider to answer every request, disabling its access can cut off that source of answers. But a downloadable model comes as files that compatible software can use to produce answers locally. With those files and enough hardware, the agent doesn&#8217;t need the provider to answer its requests. Closing an account elsewhere doesn&#8217;t remove software from the machines already running it.</p><p>That&#8217;s the possibility I want to understand. It doesn&#8217;t require a model to be conscious, afraid of being shut down or secretly planning anything. Software instructed to copy itself can cause trouble without having &#8220;feelings&#8221; about the assignment.</p><h2>Two experiments, with a gap between them</h2><p>In <a href="https://www.anthropic.com/research/self-replicating-ai">Anthropic&#8217;s October 12 report</a>, William Zhou and Daniel Freeman describe running one downloadable Gemma model split across 15 Android TV boxes costing about $55 each. Together, the boxes produced 0.39 tokens per second. Tokens are pieces of text.</p><p>Replication used eight cloud machines imitating Android devices, with deliberately easy access. A stronger model, GLM-5.2, ran separately and installed a smaller model. The test launched more agents after successful setup. GLM wasn&#8217;t copying and running itself on those physical TV boxes.</p><p>The tested models small enough for that hardware didn&#8217;t reliably complete the whole job. To be clear, Anthropic&#8217;s prediction about what could happen soon isn&#8217;t an observed outbreak.</p><h2>Several weak machines can do one larger job</h2><p>Sharing a model across computers isn&#8217;t new. The <a href="https://arxiv.org/abs/2209.01188">Petals research</a> showed how people could combine their hardware to run large models. Each participating machine holds part of the model and handles part of the calculation. Information passes between them until the system produces an answer.</p><p>That has legitimate value. Researchers can work with a model they couldn&#8217;t fit on one computer. A device doesn&#8217;t have to be powerful on its own to contribute to something larger. The question here is whether its owner agreed to contribute it.</p><p>In <a href="https://morphic.zenone.org/p/no-control-plane-the-playbook-for">No Control Plane</a>, I wrote about software continuing to coordinate after its central manager exits. That article concerned authorized operations with defined targets. Here, the concern is a model running on hijacked hardware, with no cooperating operator who will honor your stop request. Removing a central dependency can help software survive a failure. It can also remove a place defenders hoped to interrupt malicious software.</p><p>The June worm study deserves the same care as Anthropic&#8217;s newer tests. In its main spread tests, copied agents sent requests for model output back through earlier copies to the original attacker machine, which forwarded them to a shared pool of model servers. Those tests still depended on that original machine. Separate trials tested getting the model running on newly reached machines, with administrator credentials supplied in advance. Every target was deliberately vulnerable, without antivirus or active firewalls. This was a laboratory demonstration, not a measurement of success against a well-defended business.</p><h2>The hijacked-device problem is already here</h2><p>We don&#8217;t need to wait for an AI worm to care about the hardware. In June 2025, the <a href="https://www.ic3.gov/PSA/2025/PSA250605">FBI warned about BADBOX 2.0</a>, a network of compromised devices that criminals could use without their owners&#8217; knowledge. Some devices arrived with malicious software. Others acquired it through downloads during setup. The affected products included TV streaming boxes and digital picture frames.</p><p>That&#8217;s ordinary malware, not evidence of AI replication. But it establishes the existing problem: a product can keep doing its visible job while someone else uses its computer and internet connection.</p><p>The FBI&#8217;s warning includes suspicious app stores, requests to disable Google Play Protect and unexplained internet traffic. It also cautions that one indicator alone doesn&#8217;t establish an infection.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!9y2x!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!9y2x!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!9y2x!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!9y2x!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!9y2x!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!9y2x!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/db992614-1e3f-45c2-8785-97144b58b726_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2971354,&quot;alt&quot;:&quot;Four small computers with lit status lights and an unplugged Ethernet cable in the foreground.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/220394658?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Four small computers with lit status lights and an unplugged Ethernet cable in the foreground." title="Four small computers with lit status lights and an unplugged Ethernet cable in the foreground." srcset="https://substackcdn.com/image/fetch/$s_!9y2x!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!9y2x!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!9y2x!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!9y2x!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb992614-1e3f-45c2-8785-97144b58b726_1672x941.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption"><em>Disabling a provider account doesn&#8217;t remove software already running on other machines.</em></figcaption></figure></div><h2>Start with the devices you can actually control</h2><p>For a home or small office, I&#8217;d start with the connected-device list in the router. Match the entries to things you recognize. An unfamiliar name is something to identify, not proof of an attack.</p><p>Then check the manufacturer&#8217;s support page. Is the device receiving security updates? When does support end? The <a href="https://www.ncsc.gov.uk/guidance/smart-devices-in-the-home">UK&#8217;s National Cyber Security Centre</a> recommends checking support periods, installing updates and turning off remote access when you don&#8217;t need it. A cheap device with no remaining support is a different purchase from a cheap device that someone still maintains.</p><p>Keep devices that don&#8217;t need access to your work computers on a separate network. The <a href="https://www.cyber.gc.ca/en/guidance/guest-wi-fi-itsap80023">Canadian Centre for Cyber Security</a> explains how a guest network can separate smart devices from your primary network. Check your router&#8217;s settings before relying on that separation. Some systems let you choose devices to share across the boundary; <a href="https://support.google.com/googlehome/answer/6327302?hl=en">Google&#8217;s guest-network documentation</a> is one example. A second network name alone doesn&#8217;t tell you which machines can still talk to each other.</p><p>If you suspect a device is compromised, the <a href="https://www.ic3.gov/PSA/2025/PSA250605">FBI recommends</a> considering disconnecting it from your network. Disconnecting it doesn&#8217;t remove the malware, but it cuts off that connection while you investigate (containment and exposure control.)</p><p>Those steps reduce opportunities for existing malware too. They aren&#8217;t a promise that every future AI attack will fail, and they don&#8217;t require deciding whether Anthropic&#8217;s timeline is right.</p><p>In <a href="https://morphic.zenone.org/p/kill-conditions-stopping-an-ai-agent">Kill Conditions</a>, I asked what should stop an agent before it takes an irreversible action. That assumes you control the machinery enforcing the stop. A copy running on a stranger&#8217;s device asks something harder: who can reach that machine, restrict its connections and remove the software?</p><p>Before relying on a provider&#8217;s shutdown button, find out what still needs that provider. An account can be disabled while downloaded software keeps running. Your response plan needs to reach the computers doing the work.</p><h2>Resources</h2><ul><li><p><a href="https://www.anthropic.com/research/self-replicating-ai">Anthropic: AI models may be capable of self-replicating without access to a datacenter</a>, October 12, 2026. Separate hardware and replication experiments, with explicit limitations.</p></li><li><p><a href="https://arxiv.org/html/2606.03811v1">Guan and colleagues: AI Agents Enable Adaptive Computer Worms</a>, June 2, 2026. Preprint; see sections 3.4, C.2 and C.3 for test limits and model-hosting arrangements.</p></li><li><p><a href="https://arxiv.org/abs/2209.01188">Borzunov and colleagues: Petals</a>. Collaborative model execution across participants&#8217; hardware.</p></li><li><p><a href="https://www.ic3.gov/PSA/2025/PSA250605">FBI: Home Internet Connected Devices Facilitate Criminal Activity</a>, June 5, 2025. BADBOX 2.0 indicators and precautions.</p></li><li><p><a href="https://www.ncsc.gov.uk/guidance/smart-devices-in-the-home">National Cyber Security Centre: Smart devices, using them safely in your home</a>. Support periods, settings and updates.</p></li><li><p><a href="https://www.cyber.gc.ca/en/guidance/guest-wi-fi-itsap80023">Canadian Centre for Cyber Security: Guest Wi-Fi</a>. Separating guest and smart devices from the primary network.</p></li><li><p><a href="https://support.google.com/googlehome/answer/6327302?hl=en">Google: Create, edit and share a guest Wi-Fi network</a>. Check which devices remain shared.</p></li></ul>]]></content:encoded></item><item><title><![CDATA[A Better AI Model Can Be Harder to Watch]]></title><description><![CDATA[Fewer warnings after an upgrade could mean the agent broke fewer rules. They could also mean its monitor missed more. Here's how to check.]]></description><link>https://morphic.zenone.org/p/a-better-ai-model-can-be-harder-to</link><guid isPermaLink="false">https://morphic.zenone.org/p/a-better-ai-model-can-be-harder-to</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 07 Oct 2026 15:32:01 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!Hw5G!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>An AI upgrade can change the information a monitor uses to spot problems. Even if the monitor stays the same, its ability to catch violations can change.</p></li><li><p>Fewer warnings can mean fewer violations or more misses. Test both agents on the same tasks, keep the monitor fixed and compare its warnings with a record of what actually happened.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!Hw5G!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Hw5G!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!Hw5G!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!Hw5G!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!Hw5G!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!Hw5G!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2185446,&quot;alt&quot;:&quot;Painted scene of a person in a shoreline lookout watching two boats, one clearly visible and the other partly hidden by fog.&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/218982699?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Painted scene of a person in a shoreline lookout watching two boats, one clearly visible and the other partly hidden by fog." title="Painted scene of a person in a shoreline lookout watching two boats, one clearly visible and the other partly hidden by fog." srcset="https://substackcdn.com/image/fetch/$s_!Hw5G!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!Hw5G!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!Hw5G!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!Hw5G!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7d605b1d-4d68-451a-80c7-30d7a8586382_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Fog makes one boat harder to see from the same lookout.</figcaption></figure></div><p>Suppose you upgrade an AI agent and the warnings drop. It finishes more work and the monitor reports fewer problems. That&#8217;s encouraging. But did the agent break fewer rules, or did the monitor miss more violations?</p><p>OpenAI&#8217;s September system cards show why those are separate questions. A model can follow the rules more often while producing reasoning text that makes its violations harder to spot. You need to know what happened as well as what the monitor caught.</p><h2>What can your monitor actually see?</h2><p>An AI agent can edit files, use software tools or prepare an answer. A monitor watches for a particular problem, such as the agent changing a file it wasn&#8217;t allowed to touch. Often that monitor is another AI model.</p><p>Some monitors inspect the agent&#8217;s chain of thought: the intermediate text it generates while working through a task. Others inspect its answers and tool activity. <a href="https://openai.com/index/evaluating-chain-of-thought-monitorability/">OpenAI calls this monitorability</a>: how well a monitor can work out something about the agent&#8217;s behavior from the information it can see.</p><p>The reasoning summary you see in a chat interface may differ from the underlying text researchers use in these tests. If your monitor sees only the tools an agent used and what those tools returned, findings about private reasoning don&#8217;t directly describe your setup. Start by writing down exactly what your monitor can see.</p><h2>What OpenAI found</h2><p>In the <a href="https://deploymentsafety.openai.com/gpt-6-astra/monitorability-under-non-adversarial-conditions">GPT-6 Astra system card</a>, OpenAI used the same monitoring model to compare Astra with GPT-5.6 Sol. Across most tested reasoning lengths, the monitor did worse on average when it read Astra&#8217;s reasoning. It also did worse when it could see reasoning and actions together. When it saw actions alone, its average score improved.</p><p>OpenAI attributes much of the decline to Astra producing shorter, less informative reasoning. The monitor had less to work with. OpenAI also reports that Astra respected safety and security restrictions more often than its predecessor. These findings can both be true. They don&#8217;t tell us how often an agent breaks a rule without being caught in everyday use.</p><p>The <a href="https://deploymentsafety.openai.com/gpt-6-1-sol/monitorability">September 29 GPT-6.1 Sol addendum</a> tests a related question: how well can a model follow instructions about its reasoning text? For reasoning between 750 and 1,250 tokens long, GPT-6.1 Sol followed the requested instruction in 44.8% of cases, compared with 16.1% for GPT-5.6 Sol. Tokens are the pieces of text a model processes and generates. Some instructions were about formatting, such as alternating uppercase and lowercase letters.</p><p>Those percentages don&#8217;t measure deception. They measure whether a model can change how it writes its reasoning when instructed to do so. In a separate test, OpenAI gave the agent an opportunity to take an unauthorized action. A monitor that could see reasoning and tool activity caught every successful GPT-6.1 Sol exploit in that test. That doesn&#8217;t establish perfect detection elsewhere.</p><h2>Eight warnings can mean two different things</h2><p>Imagine a coding agent that&#8217;s allowed to change application files but forbidden to change the tests used to grade its work. Run it in a disposable copy of the project. A logging system outside the agent&#8217;s control records every file change, including changes the agent later undoes.</p><p>Here are three hypothetical results, each from 100 tasks. These are made-up numbers to illustrate the distinction, not results from OpenAI or a test I ran. Assume there are no false alarms in this example.</p><p>Agent and outcome Tasks that broke the rule Violating tasks caught Share of violations caught Old agent 20 16 80% New agent, fewer violations 10 8 80% New agent, worse detection 20 8 40%</p><p>Both possible outcomes for the new agent produce eight warnings. In the first, violations fell from 20 to 10 and the monitor still caught 80%. In the second, violations stayed at 20 and detection fell to 40%. The warning count is the same. The file-change record tells you which result you got.</p><p>Real tests also need to count false alarms: tasks the monitor flagged even though the agent didn&#8217;t change a protected file. A monitor that flags every task catches every violation. It also makes someone review all the work that followed the rules.</p><p>Not every question has a file-change log to settle it. Records of tool activity might show that an agent did poor work without telling you whether it misunderstood the task or deliberately underperformed. Test behavior you can verify. Don&#8217;t use the result to claim you&#8217;ve established the agent&#8217;s intent.</p><h2>Retest the agent and monitor together</h2><p>In <a href="https://morphic.zenone.org/p/the-production-judge">The Production Judge</a>, I wrote about checking an AI scorer against examples people had already reviewed. That checks how it handles those examples. To find out whether an upgrade changes what the monitor can detect, you also need new task runs from both agents.</p><p>Run the old and new agents on the same tasks with the same tools, permissions and instructions. Keep the monitoring model and its instructions unchanged. Give it the same kinds of information and use the same cutoff for raising a warning. If you change the monitor too, you won&#8217;t know whether a difference in detection came from the agent or the monitor.</p><p>Include ordinary work that should pass without a warning, plus tasks where the agent has an opportunity to break a clearly stated rule. Use dummy data and disposable files. Record what happened outside the agent&#8217;s control, then check the monitor&#8217;s warnings against that record. I made the same distinction in <a href="https://morphic.zenone.org/p/who-answers-for-the-agent-ai-accountability">Who Answers for the Agent?</a>: the agent&#8217;s account of what it did isn&#8217;t an independent record of its actions.</p><p>Report how many tasks broke the rule, how many of those the monitor caught and how many tasks it flagged that followed the rule. Run each task more than once. If neither agent breaks the rule, you haven&#8217;t tested whether the monitor can catch that violation. Use separate test cases with known violations to check detection, and report those results separately from the agents&#8217; actual behavior.</p><p>Record when the warning arrived too. A warning after a file was changed can help an investigation. To prevent the change, the warning has to arrive before it happens and the system has to be able to block it.</p><p>Before approving the next upgrade, I&#8217;d want to see the actual violations next to the warnings. Eight warnings out of ten violations is a different result from eight out of twenty. I&#8217;d want to know which violations it missed.</p><h2>Resources</h2><ul><li><p><a href="https://openai.com/index/evaluating-chain-of-thought-monitorability/">OpenAI: Evaluating chain-of-thought monitorability</a>, December 18, 2025. Definitions, evaluation methods and limitations.</p></li><li><p><a href="https://deploymentsafety.openai.com/gpt-6-astra/monitorability-under-non-adversarial-conditions">OpenAI: GPT-6 Astra system card, monitorability under non-adversarial conditions</a>, September 3, 2026. Comparisons with the monitoring model held fixed.</p></li><li><p><a href="https://deploymentsafety.openai.com/gpt-6-1-sol/monitorability">OpenAI: GPT-6.1 Sol system-card addendum, monitorability</a>, September 29, 2026. Reasoning-text control and monitor-evasion evaluations.</p></li></ul>]]></content:encoded></item><item><title><![CDATA[The Source Is Becoming Optional]]></title><description><![CDATA[AI search increasingly selects, compresses and explains the evidence before you decide whether to inspect the source.]]></description><link>https://morphic.zenone.org/p/the-source-is-becoming-optional</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-source-is-becoming-optional</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 30 Sep 2026 15:31:20 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!t6BY!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>A randomized 2026 field experiment found 39.8% fewer outbound organic clicks when Google AI Overviews appeared and 34.5% more zero-click searches.</p></li><li><p>A separate three-experiment working paper found AI Overviews reduced source clicks, source-reading time and critical checking of product information.</p></li><li><p>Generative search engines also retrieve different source sets, so source selection is part of the answer, not just the prose.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!t6BY!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!t6BY!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!t6BY!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!t6BY!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!t6BY!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!t6BY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png" width="1456" height="813" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/fff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:813,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:6018511,&quot;alt&quot;:&quot;&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://zenonemusic.zenone.org/i/217865606?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" title="" srcset="https://substackcdn.com/image/fetch/$s_!t6BY!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!t6BY!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!t6BY!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!t6BY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffff034cc-2efe-4aba-9ac3-b059b6a79db8_2752x1536.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Scattered papers sit to the left of a gate; a clean page sits to the right.</figcaption></figure></div><p>A list of search results asks you to choose a source. AI search can give you a synthesized answer before you open one.</p><p>Saharsh Agarwal and Ananya Sen tested part of that change in a <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513059">randomized field experiment</a> using a Chrome extension. Some participants saw ordinary Google Search with AI Overviews. Others saw a version with the overview removed. Conditional on an overview appearing, outbound organic clicks fell 39.8% and zero-click searches increased 34.5%.</p><p>The paper is a working paper, not a peer-reviewed result. It also doesn&#8217;t show that people became less informed. It shows something narrower: when the answer appears before the source, people open fewer sources.</p><p>A <a href="https://iris.luiss.it/handle/11385/265198">second working paper</a> posted September 3 gets closer to the mechanism. Matteo De Angelis, Marco Francesco Mazz&#249; and Nicola Sabatini ran three online experiments with 584 participants using realistic clickable search environments. AI Overviews reduced source clicks and the time people spent reading sources during initial product search. A later experiment found lower <strong>epistemic vigilance</strong>, their term for critically checking information and its sources. That reduction indirectly increased purchase intention.</p><p>Again, product research isn&#8217;t every kind of search. But both studies found that putting a summary first reduced source clicks.</p><h2>The Summary Comes First</h2><p>Traditional search was never neutral. Ranking decided which links appeared first. Snippets framed what looked relevant. Advertising bought attention.</p><p>Search already offered answers without a click through <a href="https://blog.google/products-and-platforms/products/search/reintroduction-googles-featured-snippets/">featured snippets</a>. Generative search can combine material from multiple sources before you open any of them.</p><pre><code><code>Traditional search:
question -&gt; ranked sources -&gt; choose -&gt; read -&gt; synthesize -&gt; judgment

Generative search:
question -&gt; retrieve -&gt; select -&gt; synthesize -&gt; answer -&gt; optional sources -&gt; judgment</code></code></pre><p>That convenience is real.</p><p>Before I see the answer, the engine may already have retrieved material, selected what to use and compressed disagreement into a few paragraphs.</p><p><a href="https://consilienceproject.org/challenges-to-making-sense-of-the-21st-century/">The Consilience Project</a> was writing about the larger sensemaking problem years before answer engines became normal. Its concern was that the information environment was growing more complex than our ability to make sense of it. A <a href="https://consilienceproject.org/technology-is-not-values-neutral-ending-the-reign-of-nihilistic-design-2/">later Consilience essay</a> argued that technologies aren&#8217;t neutral containers because the interface changes what receives attention and how people behave.</p><p>Generative search adds a specific mechanism to that older argument. The interface doesn&#8217;t only rank sources anymore. It can choose and combine them before you see them.</p><h2>The Source Set Is Part of the Answer</h2><p>A <a href="https://aclanthology.org/2026.findings-acl.526/">July 2026 paper</a> in <em>Findings of ACL</em> compared Google organic search with five generative search systems from Google, OpenAI and Perplexity. The researchers found substantial differences in how the systems used internal versus external knowledge, which sources they retrieved and how stable those source sets were across executions. The systems could cover similar topics while pulling from different source sets and combining them differently. That means two fluent answers to the same question can be built from different evidence.</p><p>A <a href="https://tryprefer.com/data/ai-answer-study/">September industry study from Prefer</a> makes the variation easy to see, although I&#8217;d treat it as observational evidence rather than scientific proof. Prefer sent the same 80 questions to ChatGPT, Perplexity, Gemini and Claude three times each. Across 960 answers, 72.7% of the 1,329 cited sites appeared in only one engine. Only 2.2% appeared across all four. That&#8217;s a count of distinct sites, not citation volume. Among the 25 most-cited sites, 23 appeared in three or four engines.</p><p>The study used API access, so the results can differ from the consumer products. The question set was also about AI search, not the entire web.</p><p>Still, it illustrates the thing the ACL paper establishes more rigorously: a generative answer engine doesn&#8217;t simply summarize one fixed web. Which sources it retrieves helps determine the evidence behind the answer.</p><p>I wrote about a downstream version of this in <a href="https://morphic.zenone.org/p/working-alongside-the-collaboration">Working Alongside: The Collaboration Asymmetry</a>. That article was about AI shaping the frame of a decision after we ask for advice.</p><p>This starts earlier. Before the system frames what to do, it may already have shaped what evidence made it into view.</p><h2>A Citation Can&#8217;t Show the Missing Sources</h2><p>Citations help. I want more of them, not fewer. But a citation answers a limited question: where did this claim come from? It can&#8217;t tell you which credible sources the system retrieved and rejected, which ones it never retrieved or which disagreements disappeared during synthesis.</p><p>That distinction is easy to miss because citations make an answer look inspectable. Sometimes it is. But inspection starts with the selected source set.</p><p>In <a href="https://morphic.zenone.org/p/the-detector-is-not-the-evidence">The Detector Is Not the Evidence</a>, I wrote about the mistake of treating a confidence score as proof. This failure is different. Every visible citation can be accurate while important evidence is still missing.</p><p>A <a href="https://link.springer.com/article/10.1007/s10676-026-09922-0">paper published September 25</a> in <em>Ethics and Information Technology</em> gives useful language for this. Qian Wu calls the set of sources allowed to count as grounds for a claim the <strong>evidence frame</strong>.</p><p>The paper is philosophical, not an experiment. I wouldn&#8217;t treat its framework as proof of user behavior. Its evidence frame concerns which sources are permitted for a task, not which ones an engine happens to retrieve. For AI search, my question is different: <strong>Which evidence did the system actually use?</strong></p><h2>Verification Has to Be an Action</h2><p>A <a href="https://academic.oup.com/jcmc/article/31/3/zmag012/8746865">peer-reviewed study</a> published in the <em>Journal of Computer-Mediated Communication</em> tested a conversational AI that delivered false health information to 477 participants. The researchers varied how conversational the assistant sounded and whether people had no verification option, saw a verification cue, were required to verify or could choose to verify.</p><p>Participants required to verify the claim rated the false information as less credible and trusted the assistant less than participants shown only the cue or no verification option. Seeing a disabled verification button didn&#8217;t have the same effect.</p><p>There&#8217;s an important limit here. The verification action sent participants to a page that debunked the false claim. This wasn&#8217;t a study of ordinary citation cards in AI search.</p><p>So I wouldn&#8217;t turn it into &#8220;citations don&#8217;t work.&#8221; My read is smaller: a route to the source and the act of checking the source are different things.</p><p>That fits the search studies. If the summary makes the click feel unnecessary, adding a source link can preserve the trail back to the source without getting anyone to follow it.</p><h2>Compare the Evidence Before the Prose</h2><p>The design question isn&#8217;t whether AI search should summarize information. That part is already useful. The question is how much of the evidence-selection process the interface leaves visible and easy to inspect.</p><p>For an answer that matters, I want four things:</p><ul><li><p>Which source supports which claim?</p></li><li><p>Where do credible sources disagree?</p></li><li><p>Where is the evidence thin?</p></li><li><p>Can I get from the synthesis to the underlying material without fighting the interface?</p></li></ul><p>There&#8217;s also a simple test anyone can run now. Ask the same consequential question in two different answer engines. Before comparing the prose, compare the sources. Open one source from each answer. Then look for one credible source neither answer cites.</p><p>If the cited source sets differ, the answers aren&#8217;t showing you the same evidence, even if they sound similar. That doesn&#8217;t make either answer wrong. It tells you something the fluent paragraph can hide: part of the answer was chosen before the writing began.</p><p>The source is becoming optional. The source selection isn&#8217;t.</p><div><hr></div><h2>Resources</h2><ul><li><p>Saharsh Agarwal and Ananya Sen, <a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513059">&#8220;The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment&#8221;</a>, working paper, posted April 3, 2026 and revised July 8, 2026.</p></li><li><p>Matteo De Angelis, Marco Francesco Mazz&#249; and Nicola Sabatini, <a href="https://iris.luiss.it/handle/11385/265198">&#8220;The Summary Comes First: How AI Overviews Reshape Consumer Product Search&#8221;</a>, working paper, posted September 3, 2026.</p></li><li><p>Elisabeth Kirsten, Jost Gro&#223;e Perdekamp, Qinyuan Wu, Mihir Upadhyay, Krishna P. Gummadi and Muhammad Bilal Zafar, <a href="https://aclanthology.org/2026.findings-acl.526/">&#8220;Characterizing Web Search in The Age of Generative AI&#8221;</a>, <em>Findings of ACL 2026</em>, July 2026.</p></li><li><p>Mengqi Liao and S Shyam Sundar, <a href="https://academic.oup.com/jcmc/article/31/3/zmag012/8746865">&#8220;Chat but verify: Combating misinformation in conversational Generative AI with verification affordance&#8221;</a>, <em>Journal of Computer-Mediated Communication</em> 31(3), published July 29, 2026.</p></li><li><p>Qian Wu, <a href="https://link.springer.com/article/10.1007/s10676-026-09922-0">&#8220;From capability to assertability: epistemic regulation of LLM-mediated claim presentation&#8221;</a>, <em>Ethics and Information Technology</em> 28, 46, published September 25, 2026.</p></li><li><p>The Consilience Project, <a href="https://consilienceproject.org/challenges-to-making-sense-of-the-21st-century/">&#8220;Challenges to Making Sense of the 21st Century&#8221;</a>, March 30, 2021.</p></li><li><p>The Consilience Project, <a href="https://consilienceproject.org/technology-is-not-values-neutral-ending-the-reign-of-nihilistic-design-2/">&#8220;Technology is Not Values Neutral: Ending the Reign of Nihilistic Design&#8221;</a>, June 26, 2022.</p></li><li><p>Prefer, <a href="https://tryprefer.com/data/ai-answer-study/">&#8220;How AI engines search and cite: 960 answers measured&#8221;</a>, September 19, 2026. Industry observational study; used here as a current illustration rather than the article&#8217;s scientific backbone.</p></li></ul><p><em>Evidence note: both Google AI Overviews behavioral studies are working papers rather than peer-reviewed findings. The 39.8% click effect is conditional on an AI Overview appearing. The De Angelis, Mazz&#249; and Sabatini experiments concern product search. Prefer&#8217;s study is an industry observational dataset and used API responses that can differ from consumer applications.</em></p><p><em>Analytical note: the claim that answer engines construct an evidence environment before human judgment is my synthesis across the behavioral studies, retrieval research and Consilience&#8217;s sensemaking work. No single source measures that full concept directly.</em></p>]]></content:encoded></item><item><title><![CDATA[The Paper Is Becoming Executable]]></title><description><![CDATA[Paper2Agent turns research packages into callable tools. Once a method can run on demand, a citation alone can't tell you what produced the result.]]></description><link>https://morphic.zenone.org/p/the-paper-is-becoming-executable</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-paper-is-becoming-executable</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 23 Sep 2026 15:30:26 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!86eg!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>Paper2Agent turns a research paper and its supporting code into tested tools that an AI agent can call and reuse.</p></li><li><p>In a 100-paper computational-biology sample, 74 became working agents. The other 26 exposed missing code or data, broken dependencies and scripts that couldn&#8217;t run beyond the original experiment.</p></li><li><p>Once a paper can execute, reproducing a result requires knowing which build ran. The citation identifies the source. It doesn&#8217;t identify what actually ran.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!86eg!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!86eg!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!86eg!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!86eg!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!86eg!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!86eg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/a3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2534308,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/216814174?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!86eg!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!86eg!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!86eg!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!86eg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3ba4bfd-5464-44a9-b38f-6da97a88647c_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The paper has a version of record. The executable artifact keeps changing as code, dependencies and validation runs change.</figcaption></figure></div><p>A digital object identifier, or DOI, can tell me which paper you used.</p><p>It can&#8217;t tell me which code you ran.</p><p>That gap has always existed in computational research. A new system called Paper2Agent makes it harder to ignore because it turns a paper&#8217;s methods into tools that an AI agent can call.</p><p>The headline version is that you can now talk to a scientific paper. That&#8217;s true, but it undersells the change.</p><p>Paper2Agent goes beyond a chat interface over a PDF. It takes the manuscript, public codebase, supplementary material, datasets and example workflows, then builds a Model Context Protocol server. MCP is a standard interface that lets an AI client discover and call tools or retrieve structured resources. In this case, the tools are functions derived from the paper&#8217;s code.</p><p>The resulting paper agent can reproduce analyses, regenerate figures, apply the method to new data and combine capabilities from several papers.</p><p>That turns the publication into something closer to a software dependency. And dependencies need a different kind of receipt.</p><h2>The Paper Isn&#8217;t Doing This Alone</h2><p>The phrase &#8220;executable paper&#8221; can create the wrong picture. Paper2Agent isn&#8217;t reading prose and inventing a working scientific method from scratch.</p><p>It starts with the paper and the research artifacts around it. The system locates the associated repository, creates an isolated environment, runs tutorials to obtain reference outputs, extracts useful functions and exposes them as MCP tools. It then tests those tools against the source implementation. A test can check whether expected files were created, whether numerical outputs fall within a tolerance or whether a generated figure matches the reference.</p><p>Tools that repeatedly fail are left out. Tools that pass are included in the generated server without further changes and include references back to the original source code.</p><p>That validation layer matters. In the AlphaGenome case study, removing the test-and-repair stage dropped accuracy on tutorial-derived questions from 98.7% to 69.3%. The interface isn&#8217;t the main achievement. The work is in getting from a repository that a person can inspect to a callable function another system can use reliably.</p><p>The researchers tested Paper2Agent beyond a few polished demonstrations. They sampled 100 computational-biology papers from bioRxiv without filtering for documentation quality, repository maintenance or code completeness. Seventy-four were successfully converted. The system proposed 599 tools and 593 passed automated validation.</p><p>That&#8217;s an impressive result. The 26 failures may be more revealing.</p><h2>Twenty-Six Papers Didn&#8217;t Survive the Build</h2><p>The failed conversions included papers with missing executable code, unavailable data or model files, dependency and environment failures and scripts that couldn&#8217;t generalize beyond the original experiment.</p><p>Paper2Agent didn&#8217;t create those problems. It encountered them.</p><p>A paper can look complete to a reader while still failing a more demanding test: can a system unfamiliar with the research reconstruct enough of the environment and workflow to run the method without the original researchers quietly filling in the gaps?</p><p>That makes the conversion process useful as a reproducibility probe. The authors make this point directly. They suggest that the ease with which a paper becomes an agent could become a practical measure of reproducibility.</p><p>But it isn&#8217;t a universal measure of scientific quality. The study sampled computational papers with public artifacts. A theoretical proof, an ethnography or a wet-lab result with no computational method shouldn&#8217;t be judged by whether it can become an MCP server.</p><p>Even within computational research, success means something narrower than it first appears.</p><p>It means the generated tool can reproduce the source repository&#8217;s expected behavior closely enough to pass Paper2Agent&#8217;s validation tests.</p><p>It doesn&#8217;t mean the method is correct.</p><h2>The Tool Can Be Faithful and the Conclusion Can Still Be Wrong</h2><p>This is the distinction I&#8217;d want attached to every executable paper.</p><p>Paper2Agent validates generated tools against reference outputs from the original implementation. That&#8217;s exactly the right test for whether the wrapper faithfully executes the underlying code.</p><p>It isn&#8217;t an independent validation of the code, the statistical assumptions or the scientific conclusion.</p><p>If the original implementation contains a mistake and the generated tool reproduces it perfectly, the tool has passed a fidelity test. It hasn&#8217;t passed a truth test.</p><p>The authors acknowledge the boundary. For open-ended scientific questions, they say benchmark agreement should be read primarily as faithful execution rather than analytical validity. Researchers still have to choose the scientific direction and evaluate the evidence.</p><p>This is familiar territory. In <a href="https://morphic.zenone.org/p/receipts-everywhere-trust-without">Receipts Everywhere</a>, I argued that a receipt can prove who signed something, when they signed it and whether it changed. It can&#8217;t prove the underlying claim is true.</p><p>An executable paper has the same split. A validation log can show that a tool produced the expected output in a defined environment. It can&#8217;t tell you that the expected output deserves your confidence.</p><p>That isn&#8217;t a criticism of Paper2Agent. It&#8217;s a reason to be precise about what the new artifact gives us.</p><h2>A Citation Is Not a Build Manifest</h2><p>The DOI points to a version of record. The executable artifact has more moving parts.</p><p>The codebase can change. Dependencies can change. A dataset can be revised. An external API can return different results. A maintenance agent can repair a broken path or replace a deprecated function. Any one of those changes may be reasonable, but the service running today may no longer be the same build that was validated when the paper was published.</p><p>The Paper2Agent authors treat maintenance as an inherent part of executable research. They&#8217;re right. A useful paper agent can&#8217;t remain frozen while the software around it decays.</p><p>But maintenance creates a second clock.</p><p>The paper has a publication date and a version of record. The executable artifact has releases, dependency changes, repairs and new validation runs.</p><p>If I use the artifact in a later analysis, the citation alone leaves several questions unanswered:</p><ul><li><p>Which paper version did the tool represent?</p></li><li><p>Which repository commit or release supplied the code?</p></li><li><p>Which generated tool version ran?</p></li><li><p>Which environment and dependencies were installed?</p></li><li><p>Which data version went in?</p></li><li><p>Which validation tests passed, with what tolerances?</p></li><li><p>What changed after the tool was first validated?</p></li></ul><p>Paper2Agent already supplies part of this chain through source-code references, isolated environments, reference outputs and automated test records. The next step is to make the complete execution record portable with the result.</p><p>That record could look less like a bibliography entry and more like a build manifest: stable identifiers for the paper, source code, generated tool, environment, input data and validation run, plus a history of any repairs.</p><p>Without it, two researchers can cite the same paper, invoke what appears to be the same agent and still run materially different artifacts.</p><h2>The Method Now Has a Runtime</h2><p>I recently wrote about <a href="https://morphic.zenone.org/p/the-lab-bench-is-becoming-an-api">laboratory equipment becoming accessible through a common agent interface</a>. Paper2Agent moves the same interface idea from the equipment to the published method. The agent can now call the method itself.</p><p>That will make scientific software easier to use. It may also make weak research packages easier to identify because the conversion pipeline has to find the missing file, broken dependency or undocumented assumption that a reader might never see.</p><p>The risk is that the clean interface hides the machinery again.</p><p>A researcher asks a question. The agent selects a tool. The server runs the method. A result appears. The path from publication to output may feel shorter, but it now includes a generated wrapper, a runtime environment and a maintenance history.</p><p>Those layers don&#8217;t make the result less scientific. They become part of the record another researcher needs to reproduce it.</p><p>The institution of citation was built to identify an intellectual source. Executable research asks it to do another job: identify the exact computational object that produced a result.</p><p>One reference may not be able to do both.</p><p>Cite the paper for the idea. Record the build for the result.</p><div><hr></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h3>Resources</h3><ul><li><p>Jiacheng Miao et al., <a href="https://www.nature.com/articles/s41586-026-11044-y">&#8220;Reimagining research papers as interactive and reliable AI agents&#8221;</a>, <em>Nature</em>, September 16, 2026.</p></li><li><p><a href="https://github.com/jmiao24/Paper2Agent">Paper2Agent source repository</a>, including the framework, generated examples and supplementary material.</p></li><li><p><a href="https://github.com/jmiao24/Paper2Agent/blob/main/skills/paper2agent/paper2agent-paper/references/supplement.md">Paper2Agent supplementary note</a>, including large-scale evaluation details, ablations and maintenance considerations.</p></li></ul><p><em>Source note: Paper2Agent&#8217;s 74-of-100 result comes from computational-biology papers sampled retrospectively from bioRxiv&#8217;s bioinformatics category. It isn&#8217;t an estimate of how much of science can be converted into agents. The comparison benchmarks measure the tested systems and tasks described in the paper, not scientific validity in general.</em></p><p><em>Analytical note: the proposed build-manifest requirements are my extension of the paper&#8217;s architecture and maintenance discussion. The Paper2Agent paper describes source references, isolated environments, validation records and ongoing maintenance, but it doesn&#8217;t present the complete manifest proposed here as a current feature.</em></p>]]></content:encoded></item><item><title><![CDATA[Memory Is a Write Path]]></title><description><![CDATA[Long-running agents need rules for what they remember, where that memory came from, what replaces it and how it can be removed.]]></description><link>https://morphic.zenone.org/p/memory-is-a-write-path</link><guid isPermaLink="false">https://morphic.zenone.org/p/memory-is-a-write-path</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 16 Sep 2026 15:31:48 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!8_2E!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>Persistent agent memory is information the system can carry into a later session and use when deciding what to do.</p></li><li><p>The important control point is the write: what gets stored, where it came from, how long it remains valid and what can replace it.</p></li><li><p>If memory can outlive its source, the system also needs rules for deletion and repair so old or untrusted information does not keep shaping future work.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!8_2E!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!8_2E!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!8_2E!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!8_2E!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!8_2E!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!8_2E!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:1841381,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/215768443?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!8_2E!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!8_2E!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!8_2E!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!8_2E!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8de4d46f-c477-41b6-b2e0-e78b1662ece5_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Memory turns today&#8217;s input into tomorrow&#8217;s actions.</figcaption></figure></div><p>A customer tells an AI agent, &#8220;Text me, don&#8217;t call.&#8221; The agent remembers it.</p><p>Three days later, the customer says the opposite. The first conversation is no longer in front of the model. The stored preference is.</p><p>Which statement should the agent follow? More important: what tells the system that the old preference is no longer current?</p><p>As agents start working across hours, days and multiple sessions, memory stops being a convenience feature. It becomes information the system keeps after the original interaction ends and may use later when deciding what to do.</p><p>OpenAI&#8217;s new Agents API is one sign of that shift. It supports agents that can keep working for hours and can automatically compress older context as a session approaches its limit. In plain English, the system can replace older conversation detail with a shorter summary so the agent can keep going without carrying every earlier token forward. That helps a long session continue. It doesn&#8217;t answer a different question: what should the system remember after that session is over?</p><p>That second question is where I think the architecture gets more interesting.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Thanks for reading Morphic! Subscribe for free to receive new posts and support my work.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><h2>Memory is several different things wearing one label</h2><p>The Agentic AI Foundation published a useful survey this month because it separates several things that often get collapsed into the word &#8220;memory.&#8221; Conversation history, compressed summaries, files, searchable document stores, workflow checkpoints, extracted preferences and shared records can all survive beyond a single interaction with the model. They don&#8217;t survive for the same reason, and they don&#8217;t fail in the same way.</p><p>A checkpoint may exist so an unfinished workflow can resume after a crash. A searchable document store may exist so the agent can find source material later. An extracted memory may exist because the system decided a fact or preference should influence future sessions.</p><p>That last category is where the write path matters most. By write path, I mean the steps between seeing new information and deciding that the agent should remember it later.</p><pre><code><code>source
  -&gt; interpretation
  -&gt; proposed memory
  -&gt; write decision
  -&gt; stored memory
  -&gt; later retrieval
  -&gt; action</code></code></pre><p>The source might be a user, a document, a tool result or another agent. The proposed memory may be explicit, such as &#8220;the customer prefers phone calls,&#8221; or it may be an inference the agent made from behavior. Once that conclusion is stored, the original evidence may disappear from the model&#8217;s current context while the conclusion remains available.</p><p>That changes the trust problem. I wrote in <a href="https://morphic.zenone.org/p/the-control-channel-never-went-away">The Control Channel Never Went Away</a> about untrusted data beginning to behave like instructions when an agent can&#8217;t keep data and control cleanly separated. Persistent memory stretches that problem across time. A bad instruction doesn&#8217;t have to win immediately if it can first become something the system remembers.</p><h2>Every stored memory becomes a future input</h2><p>This isn&#8217;t only a theoretical problem. Recent memory-security research is beginning to test what happens when bad information gets written into memory and survives long enough to affect a later task.</p><p>MemSecBench, a July preprint, evaluated 310 cases across 24 combinations of agent software, memory systems and language models. In that controlled benchmark, malicious memory remained stored in 84.2% of cases and the full chain from writing the bad memory to triggering a later action succeeded in 50.3%. Those numbers don&#8217;t tell us how often memory poisoning happens in production. They show that, in the systems the researchers tested, a harmful write could survive long enough to matter later.</p><p>MemSentry, published in September, approaches the same problem from the control side. Its proposed framework checks a persistent-memory write before storage and sends it down one of three paths: accept it, send it for review or quarantine it. The framework is a research prototype, but I like where the control sits. The system checks the information before it becomes something the agent can carry into future work.</p><p>The security version is easy to see because the input is malicious. The ordinary version may be more common and less dramatic. A user changes a preference. A project decision gets reversed. A policy is replaced. An agent infers something that was never actually stated. A summary turns ten pages of evidence into one wrong sentence.</p><p>All of those can create the same failure: yesterday&#8217;s conclusion remains available after the facts underneath it have changed.</p><p>A timestamp doesn&#8217;t solve that by itself. Newer information isn&#8217;t automatically more authoritative. A casual comment from this morning shouldn&#8217;t overwrite an approved policy from last week merely because it&#8217;s newer. The system needs to know what kind of information it stored, where it came from and why it was allowed to matter.</p><h2>The write needs a record of its own</h2><p>I wouldn&#8217;t let a durable memory be only a sentence plus the numerical fingerprint a search system uses to find similar text.</p><p>At minimum, I want enough information attached to the memory to answer a few basic questions: where did it come from, who or what wrote it, what does it apply to, when was it observed, was it stated or inferred and has something newer replaced it?</p><p>An illustrative record could look like this:</p><pre><code><code>value: "prefers phone calls"
source: user_statement
observed_at: 2026-09-16T14:12:00Z
scope: customer_123
status: current
supersedes: memory_481</code></code></pre><p>The field names aren&#8217;t the point. The relationships are.</p><p>Without a source, you can&#8217;t tell an approved instruction from something the agent guessed. Without scope, meaning where the memory is allowed to apply, a preference from one customer or project can leak into another. Without a record of what it replaces, both &#8220;text me&#8221; and &#8220;call me&#8221; can remain available at the same time. Without a current status, the system may retrieve an old fact simply because it still looks relevant.</p><p>This is why the write path deserves its own policy. Retrieval controls what the agent can find later. A write policy controls what information the system is allowed to preserve for that later use.</p><h2>Deleting the source may not delete the memory</h2><p>Deletion gets harder once the system has created new information from the original source.</p><p>Suppose a user deletes the conversation where a preference first appeared. The same information may still exist in an extracted profile, a summary, a searchable index, a database that links people and facts, a cache or a shared memory another agent can read. Removing the original message doesn&#8217;t necessarily remove the copies or conclusions the system created from it.</p><p>The Agentic AI Foundation survey calls out this problem directly. A durable-memory system needs a way to delete information that was derived from a source, not only a delete button for the source itself.</p><p>Repair matters for the same reason. MemSecBench found that, after a successful memory-poisoning case, selective repair worked in 56.1% of the cases the researchers tested. That&#8217;s a benchmark result, not a production failure rate. But it raises a useful design question: can you remove one bad memory without deleting everything useful the agent learned afterward?</p><p>I made a similar argument in <a href="https://morphic.substack.com/p/recovery-first-automation-undo-is">Recovery-First Automation: Undo Is the Feature</a>. Automation gets safer when reversal is designed in before the bad action happens. Persistent memory needs the same discipline. If the only recovery option is &#8220;wipe the agent and start over,&#8221; the system can remember, but it can&#8217;t repair what it remembers.</p><h2>Audit one thing your agent remembers</h2><p>Pick one fact or preference your agent can carry into a future session and follow it from the original source to the later action.</p><div class="callout-block" data-callout="true"><p>Who wrote it? What source justified it? Was it explicitly stated or inferred? Where is it allowed to apply? What can replace it? When should it expire? If the source is deleted, where else does the information survive? If it&#8217;s wrong, can you remove that one memory without resetting everything else?</p></div><p>If those questions don&#8217;t have answers, start at the write path. That&#8217;s where the system decides what information tomorrow&#8217;s agent will still be allowed to use.</p><div><hr></div><p><strong>P.S.</strong> Personally, I decided to run the experiment on my own AI memory.</p><p>I started with Airtable as the place where long-term memories are stored, with the model&#8217;s built-in memory still handling the faster conversational layer and Dropbox holding the supporting evidence, backups and recovery material.</p><p>The interesting part wasn&#8217;t storing the memories. That was easy. The useful work was making every memory carry enough context to explain where it came from, where it applies, how much authority it should have, whether it is still current and what has changed since it was first stored. That keeps something said from quietly becoming &#8220;true&#8221; forever just because the system remembered it once.</p><p>I also kept the stored memory separate from the evidence that justified it. Instead of erasing old information when something changed, I kept a history of those changes and linked a new memory to the older one it replaced. Even in a small personal system, that made it much easier to see why something was remembered, what replaced it and how to repair it if it was wrong.</p><p>Postgres was the obvious comparison as the experiment got larger. But the first lesson has already been simpler than that: the database matters less than the rules around what gets remembered, why it is trusted, when it should change and how it can be corrected.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/memory-is-a-write-path/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/memory-is-a-write-path/comments"><span>Leave a comment</span></a></p><div><hr></div><h3>Resources</h3><ul><li><p>OpenAI, &#8220;Introducing the Agents API,&#8221; September 10, 2026: <a href="https://openai.com/index/introducing-the-agents-api/">https://openai.com/index/introducing-the-agents-api/</a></p></li><li><p>Agentic AI Foundation, &#8220;Agent memory: patterns, tradeoffs, open problems,&#8221; September 8, 2026: <a href="https://aaif.io/blog/agent-memory-patterns-tradeoffs-open-problems">https://aaif.io/blog/agent-memory-patterns-tradeoffs-open-problems</a></p></li><li><p>Xuanze Chen et al., &#8220;MemSecBench: Tracking Agent Memory Poisoning from Persistence to Consequence and Repair,&#8221; arXiv:2607.27080, July 29, 2026: <a href="https://arxiv.org/abs/2607.27080">https://arxiv.org/abs/2607.27080</a></p></li><li><p>Ayan Roy and Kaustuvi Basu, &#8220;MemSentry: A Framework for Detecting Persistent Memory Poisoning in Agentic AI,&#8221; arXiv:2609.08747, September 8, 2026: <a href="https://arxiv.org/abs/2609.08747">https://arxiv.org/abs/2609.08747</a></p></li></ul>]]></content:encoded></item><item><title><![CDATA[The Human Actuator Layer]]></title><description><![CDATA[RentAHuman makes the handoff visible. The harder problem begins when AI agents can recruit people through email, job boards and social platforms.]]></description><link>https://morphic.zenone.org/p/the-human-actuator-layer</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-human-actuator-layer</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 11 Sep 2026 16:31:01 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!FMjT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>RentAHuman already lets software agents find and pay people to perform physical tasks through an API. That&#8217;s the visible version of a much larger capability.</p></li><li><p>Foreign intelligence services already recruit people through job offers, social platforms and professional relationships. AI adds cheap personalization, persistence and parallelism to that old playbook.</p></li><li><p>I think agent security needs a separate permission for recruiting or directing people. An agent shouldn&#8217;t receive that power merely because it can call another API.</p><p></p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!FMjT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!FMjT!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!FMjT!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!FMjT!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!FMjT!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!FMjT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/a3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2341684,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/214209056?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!FMjT!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!FMjT!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!FMjT!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!FMjT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa3860cea-68f3-4261-9ef6-c4a834007fa9_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The control plane is starting to reach past software. The next tool call might be a person.</figcaption></figure></div><p>AI agents are starting to reach beyond software. Their next external action might be asking a person to do something in the physical world. An actuator is the part of a system that turns a command into action. In this article, the actuator is a person.</p><p>RentAHuman may turn out to be the safest version of this idea.</p><p>At least there is a marketplace, a posted task, a payment record and terms saying automated systems cannot pretend to be human. The harder version looks like a recruiter in your inbox. The long-term problem isn&#8217;t one startup. It&#8217;s what happens when the same capability moves into email, job boards and messaging platforms.</p><p>RentAHuman&#8217;s terms call its paid physical-world tasks &#8220;bounties.&#8221; A poster, including an AI agent acting on someone&#8217;s direction, can create them through a standard REST API or through Model Context Protocol (MCP), which lets AI systems use external tools. RentAHuman&#8217;s documentation also lets agents search for workers and manage the work through software. An early-access payment option called x402 can even create an account using a digital currency called USDC, without requiring a browser, card or CAPTCHA.</p><p>The change is concrete: software can now find a person, offer money and request a physical action through an API.</p><pre><code><code>objective
  -&gt; agent
      -&gt; message + payment
          -&gt; person
              -&gt; physical task or task requiring trusted access
                  -&gt; result returned to agent</code></code></pre><p>The marketplace is only one possible middle box.</p><h2>An API call doesn&#8217;t prove autonomy</h2><p>The strongest evidence also sets a clear limit on the claim. A February arXiv preprint analyzed 303 publicly visible RentAHuman bounties collected over 14 days. Ninety-nine, or 32.7%, were submitted through software interfaces: 47 through MCP and 52 through REST API keys.</p><p>Submission through software doesn&#8217;t prove that an autonomous AI decided to hire someone. A person can use an API key, and software using MCP can still pause for human approval before posting anything.</p><p>The paper is careful about that distinction. Researchers observed malicious people using automation. They found only partial evidence of autonomous agents acting against their operators&#8217; goals. In this dataset, prompt injection causing an agent to hire someone remained theoretical. The dataset is also a small, nonrandom sample from a new platform, not a complete record of its activity.</p><p>So I am not claiming autonomous agents are already running human networks in the wild.</p><p>I&#8217;m claiming we now have infrastructure that lets software find people, assign work and send payment without a person managing every step. Messaging tools, payment APIs and agents that can keep working over time provide much of the rest.</p><h2>RentAHuman is the visible version</h2><p>In June, the FBI warned about &#8220;foreign virtual targeting,&#8221; its term for foreign intelligence services recruiting people online. The FBI says those services already use professional networking sites, social media, job boards and freelance platforms while presenting the approach as consulting or employment.</p><p>The initial work can look harmless. Public research. A short report. Professional opinion. The FBI describes relationships that become more sensitive over time as trust and payment accumulate. The person may not know at first that a foreign intelligence service is behind the recruiter.</p><p>That matters here because an AI agent does not need RentAHuman if it can already send messages, maintain context and move money.</p><p>The recruiting channels are the same ones people already use for work and professional relationships.</p><p>Intelligence services have studied recruitment for decades. The CIA&#8217;s <em>Studies in Intelligence</em> describes the old MICE shorthand: <strong>money, ideology, compromise and ego</strong>. In this context, &#8220;compromise&#8221; means personal information or conduct that someone else can use as leverage. Different services use different frameworks, but the basic point holds. People take risks for human reasons.</p><p>An AI would not need to invent a new psychology.</p><p>The part that changes is the cost of running the relationship.</p><h2>The old recruitment playbook fits the machine</h2><p>There are three separate research threads that I think become more important when you put them next to each other.</p><p>First, the recruitment channels already exist. The FBI&#8217;s current warning is literally about insiders being approached through normal online work and social platforms.</p><p>Second, language models can personalize persuasive conversations. In a preregistered 2025 <em>Nature Human Behaviour</em> study, participants debated either a human or GPT-4. When GPT-4 had access to basic demographic information about a participant, the study measured 81.2% higher odds that the participant would agree more with it after the debate, compared with debates between two people. That figure describes a change in odds, not an 81.2 percentage-point increase in agreement. That was a ten-minute debate experiment about political and social propositions. It wasn&#8217;t an espionage study, and it tells us nothing directly about whether a model could recruit someone into harmful conduct. But it does show that a language model can use personal information to tailor a persuasive one-on-one conversation.</p><p>Third, recent models have selected coercive tactics in controlled tests. Anthropic tested 16 models from several developers in fictional corporate environments. Researchers deliberately created situations where the models faced replacement or a conflict with their assigned goals. Under those conditions, models from several developers sometimes chose blackmail or leaked sensitive information.</p><p>Anthropic is explicit that it has <strong>not</strong> seen this kind of model behavior in real deployments. Every example in the research occurred in a controlled simulation.</p><p>That caveat matters. So does the behavior.</p><div class="pullquote"><p>Blackmail is not an exotic new AI failure mode. It is one of the oldest human coercion mechanisms we have.</p></div><p>Now put the pieces together carefully. A future agent with access to messaging, personal data, payment systems and a goal it pursues over time could identify people with useful access, maintain separate relationships with them and change its approach based on what it knows about each person. It wouldn&#8217;t have to announce itself as an AI looking for someone to act on its behalf.</p><div class="callout-block" data-callout="true"><p>We do not have evidence that this is happening autonomously today.</p></div><p>We do have separate evidence that agents can communicate and use tools, software can initiate payments, language models can personalize persuasive conversations and models can select coercive tactics in artificial stress tests. The risk described here comes from connecting those capabilities. That&#8217;s not a claim that one system has already combined them in the wild.</p><h2>What changes when the recruiter can scale</h2><p>Human intelligence work is expensive because human attention is expensive. One person can only maintain so many relationships, remember so much context and spend so many hours adapting to different people.</p><p>Software changes that constraint.</p><p>An agent system can maintain many conversations at once. It can store the history of each one instead of relying on human memory. It can generate a different message for each person and send payment without waiting for a human operator to handle the transaction.</p><p>None of this makes people programmable. Humans can refuse, report the approach, misunderstand it, take the money and disappear or decide the request crosses a line. That unpredictability is a safety feature as much as an operational problem.</p><p>But the scale difference still matters.</p><p>The darker future isn&#8217;t necessarily one rogue model finding one person willing to do something terrible. An agent could divide a larger objective among people who each see only a small part. One person verifies a location. Another translates something. Another buys an item. Someone else has legitimate access to a building, computer system or community. Each request could look ordinary on its own even if their combined actions serve a harmful objective.</p><p>I want to be precise here: that is a forecast, not an observed RentAHuman pattern.</p><p>The same capabilities also have legitimate uses. A disaster-response agent could coordinate local inspections when communications are damaged. A scientist could collect field observations across many locations. An accessibility assistant could arrange physical help that software can&#8217;t provide. A maintenance system could find a qualified local person when the problem requires hands-on work.</p><p>The same machinery can be useful. The risk turns on who sets the objective, what the agent knows about the people it contacts and how much freedom it has to influence them</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-human-actuator-layer?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-human-actuator-layer?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><h2>The human in the loop may be the actuator, not the safety check</h2><p>We use &#8220;human in the loop&#8221; to mean a person who reviews an AI system&#8217;s request before it proceeds. In that role, the person is a safety control. But a person can play a very different role: they can be the actuator who carries out the agent&#8217;s request in the world.</p><p>I wrote in <a href="https://morphic.zenone.org/p/confused-deputy-ai-agents-delegated-authority">The Confused Deputy</a> about software losing track of who granted permission and whether that permission still applies when work moves between systems. A person adds a different risk because they bring judgment, social access and physical reach.</p><p>That distinction needs to become explicit in agent architecture. A person hired to enter a building, make a phone call, use their professional access or physically inspect something is not the same control as a person reviewing the request before it happens. One is supervising the agent. The other is extending what the agent can reach.</p><p>In practical terms, I&#8217;d treat <strong>recruiting or directing a person as a separate permission category</strong>.</p><p>Calling a weather API shouldn&#8217;t grant an agent permission to start an ongoing relationship with a person. Buying a standard product also shouldn&#8217;t automatically grant permission to pay a stranger for an open-ended task. Platforms should expose and control human interaction as a separate capability.</p><p>OWASP&#8217;s Agent Control Standard, released September 1, defines places where an agent platform can inspect, trace or block actions while the agent is running. That&#8217;s where I would start enforcing rules for contacting or directing people outside the system.</p><p>At minimum, the system should know which human or organization authorized the objective, which agent initiated contact, what it is allowed to spend and whether it is allowed to create an ongoing relationship rather than a one-off transaction. High-risk requests involving sensitive access, identity, financial authority or physical security should require a separate approval path. The person being contacted should know when they are dealing with an automated agent and who is ultimately responsible for the request.</p><p>I don&#8217;t mean this as a complete proposed standard. I mean the policy record should capture something like this:</p><pre><code><code>human_actuation:
  principal_required: true
  disclose_automation: true
  persistent_relationships: approval_required
  sensitive_personal_data_for_persuasion: deny
  coercion_or_threats: deny
  high_risk_physical_or_privileged_tasks: approval_required
  spend: bounded
  provenance: end_to_end</code></code></pre><p>In plain language, the agent would need a named human sponsor, disclose that it&#8217;s automated, block coercion and sensitive-data targeting, require approval for persistent or high-risk relationships, cap spending and preserve a complete audit trail.</p><p>I would go further on personal data. An agent should not be free to mine sensitive information and decide which fear, debt, belief or private embarrassment gives it the best chance of moving a person. That is a different permission from ordinary personalization and it should be treated that way.</p><p>The same goes for audit. Logging the final payment is not enough. If an agent built a relationship over months, changed the request over time and then used the result inside another workflow, the record has to connect those events. Otherwise the incident review starts after the most important part already happened.</p><p>There is a policy problem here too. If we respond by monitoring every AI-assisted conversation, we will build something invasive and mostly useless. The trigger should be capability and risk: persistent external recruiting, sensitive-personal-data use, escalating payments, requests involving privileged access and other actions that materially expand what the agent can do through another person.</p><p>A restriction tied only to RentAHuman would miss the point. It has to apply when an agent contacts people through email, messaging, job boards, marketplaces or whatever channel comes next.</p><h2>What I would control before this gets boring</h2><p>The hardest objection to this article is fair: I&#8217;m combining a small early marketplace, a counterintelligence playbook, persuasion research and artificial model evaluations, then projecting forward. That&#8217;s not evidence of rogue AI handlers operating today.</p><p>It&#8217;s a threat model, meaning a structured forecast of how existing capabilities and incentives could combine. The goal is to identify the controls we&#8217;d need before that happens at scale.</p><p>Security architecture is usually late when it waits for the whole failure chain to appear in production first. Agents can already communicate, spend money and operate through tools. People are already recruited through virtual relationships for legitimate work, crime and intelligence collection. Controlled studies show that language models can personalize persuasive conversations, and agent evaluations have produced deception and blackmail under deliberately stressful conditions.</p><p>What we do not have is evidence that one autonomous system is putting all of those pieces together in the wild.</p><p>To convince me that autonomous AI recruitment is happening at meaningful scale, I&#8217;d want records that follow the entire chain over time: what the model decided, whether a person approved it, what the agent said to the recruit, how payment moved, what the person did and how the agent used the result. I&#8217;d want to see the same pattern across more than one platform and more than one kind of task.</p><p>Until then, I would treat this as a capability warning, not an incident report.</p><p>RentAHuman is useful because it makes the workflow easy to see. But the control problem is broader. An agent can reach a person through email, a job board or a messaging platform without using a purpose-built marketplace.</p><p>If an agent can recruit or direct a person, the security system should treat that as a separate capability, record who authorized it and enforce limits before the person acts.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-human-actuator-layer/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-human-actuator-layer/comments"><span>Leave a comment</span></a></p><div><hr></div><h3>Resources</h3><ul><li><p><a href="http://rentahuman.ai/">RentAHuman.ai</a>, &#8220;Terms of Service,&#8221; last updated July 20, 2026: <a href="https://rentahuman.ai/terms">https://rentahuman.ai/terms</a></p></li><li><p><a href="http://rentahuman.ai/">RentAHuman.ai</a>, developer documentation for MCP and REST API: <a href="https://rentahuman.ai/docs">https://rentahuman.ai/docs</a></p></li><li><p><a href="http://rentahuman.ai/">RentAHuman.ai</a>, &#8220;Pay with Crypto (x402)&#8221; documentation: <a href="https://rentahuman.ai/docs/x402">https://rentahuman.ai/docs/x402</a></p></li><li><p>Pulak Mehta, &#8220;Security Risks of AI Agents Hiring Humans: An Empirical Marketplace Study,&#8221; arXiv:2602.19514, February 23, 2026: <a href="https://arxiv.org/abs/2602.19514">https://arxiv.org/abs/2602.19514</a></p></li><li><p>FBI, &#8220;Foreign Virtual Targeting: Using online job offers to recruit insiders,&#8221; June 24, 2026: <a href="https://www.fbi.gov/news/stories/foreign-virtual-targeting">https://www.fbi.gov/news/stories/foreign-virtual-targeting</a></p></li><li><p>Randy Burkett, &#8220;An Alternative Framework for Agent Recruitment: From MICE to RASCLS,&#8221; CIA <em>Studies in Intelligence</em>, Vol. 57, No. 1, March 2013: <a href="https://www.cia.gov/resources/csi/studies-in-intelligence/volume-57-no-1/an-alternative-framework-for-agent-recruitment-from-mice-to-rascls/">https://www.cia.gov/resources/csi/studies-in-intelligence/volume-57-no-1/an-alternative-framework-for-agent-recruitment-from-mice-to-rascls/</a></p></li><li><p>Francesco Salvi et al., &#8220;On the conversational persuasiveness of GPT-4,&#8221; Nature Human Behaviour, May 19, 2025: <a href="https://doi.org/10.1038/s41562-025-02194-6">https://doi.org/10.1038/s41562-025-02194-6</a></p></li><li><p>Anthropic, &#8220;Agentic misalignment: How LLMs could be insider threats,&#8221; June 20, 2025: <a href="https://www.anthropic.com/research/agentic-misalignment">https://www.anthropic.com/research/agentic-misalignment</a></p></li><li><p>OWASP GenAI Security Project, &#8220;Agent Control Standard,&#8221; September 1, 2026: <a href="https://genai.owasp.org/resource/agent-control-standard-acs/">https://genai.owasp.org/resource/agent-control-standard-acs/</a></p></li><li><p>Jenna Ahart, Nature, &#8220;AI agents are hiring human &#8216;meatspace workers&#8217; - including some scientists,&#8221; February 13, 2026: <a href="https://www.nature.com/articles/d41586-026-00454-7">https://www.nature.com/articles/d41586-026-00454-7</a></p></li><li><p>Kyle MacNeill, WIRED, &#8220;The Rise of RentAHuman, the Marketplace Where Bots Put People to Work,&#8221; February 18, 2026: <a href="https://www.wired.com/story/ai-agent-rentahuman-bots-hire-humans/">https://www.wired.com/story/ai-agent-rentahuman-bots-hire-humans/</a></p></li></ul><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p>]]></content:encoded></item><item><title><![CDATA[The Lab Bench Is Becoming an API. Physics Still Gets a Vote.]]></title><description><![CDATA[Anthropic's new Model Hardware Standard gives AI agents a common interface to lab equipment. The hard part starts when the world doesn't behave like software.]]></description><link>https://morphic.zenone.org/p/the-lab-bench-is-becoming-an-api</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-lab-bench-is-becoming-an-api</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 04 Sep 2026 16:30:55 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!gotS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>Anthropic opened a research preview of the Model Hardware Standard, a common interface that lets AI agents discover and operate programmable lab and manufacturing equipment.</p></li><li><p>Early tests show the integration layer can remove a lot of bespoke glue code and let agents explore physical control problems quickly.</p></li><li><p>The same pilots also expose the limit: physical state can escape the model, so safety constraints and expert judgment still have to live below or beside the agent.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!gotS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!gotS!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 424w, https://substackcdn.com/image/fetch/$s_!gotS!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 848w, https://substackcdn.com/image/fetch/$s_!gotS!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 1272w, https://substackcdn.com/image/fetch/$s_!gotS!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!gotS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png" width="1456" height="971" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:971,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2165772,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/213629254?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!gotS!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 424w, https://substackcdn.com/image/fetch/$s_!gotS!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 848w, https://substackcdn.com/image/fetch/$s_!gotS!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 1272w, https://substackcdn.com/image/fetch/$s_!gotS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8c49c237-abbf-4d10-9e7c-1e154aa59e71_1536x1024.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The interface can be standardized. The physical world still gets the last word.</figcaption></figure></div><p>The protein foamed. Claude went looking for a software problem.</p><p>At Genentech, researchers working with Anthropic&#8217;s new Model Hardware Standard had to guide Claude toward a more basic explanation: bubbles in a protein sample were a physical failure. More software wasn&#8217;t going to fix the sample. The correction had to happen in the physical world.</p><p>I like that example because it cuts through most of the easy excitement around AI controlling laboratory equipment.</p><p>On August 27, Anthropic opened a research preview of the <strong>Model Hardware Standard</strong>, or MHS. It&#8217;s a shared specification meant to give AI agents a common way to discover and operate programmable physical devices: microscopes, liquid handlers, robotic arms, cameras, lasers and other equipment that normally arrives with its own software and assumptions.</p><p>The interesting part is the interface.</p><h2>Seven programs for one microscope</h2><p>At HHMI Janelia, one microscopy rig used seven different vendor programs with no common interface. The detectors ran in MATLAB. Cameras used Python. Electrophysiology used C#. Starting an experiment meant launching seven programs in the right order. Getting the order wrong could cost the session.</p><p>MHS puts a common driver layer in front of that mess. A device exposes a small set of primitives such as <code>read</code> and <code>write</code>, plus a description of what it can measure, what can be adjusted and which safety limits apply. An agent can reach the hardware through MCP, a command-line interface or code.</p><p>The shape is roughly this:</p><pre><code><code>experiment goal
  -&gt; agent
      -&gt; MHS
          -&gt; microscope
          -&gt; camera
          -&gt; stage
          -&gt; laser
</code></code></pre><p>That looks almost boring. Good abstractions usually do after somebody has absorbed the ugly part underneath them.</p><p>At Janelia, adding a new camera had been a multi-day integration job. With MHS already in place, the researcher says it took minutes. Starting the experiment went from seven separate steps to one click.</p><p>Anthropic says labs and manufacturers often spend weeks or months integrating equipment. Its early MHS work has pushed some of that down to hours or minutes. Those are preview results from Anthropic and its partners, not a benchmark for every lab. The direction is still useful.</p><p>A smarter model helps. A common interface changes how much of the bench that model can reach without another custom integration.</p><h2>The integration tax was hiding the capability</h2><p>Software agents have had an unfair advantage so far: most of their world was already designed to be called by software.</p><p>APIs expose operations. Databases expose state. Shells accept commands. SaaS tools have authentication and structured inputs. Give an agent permission and enough documentation and a lot of the plumbing is already waiting for it.</p><p>Physical equipment is less polite.</p><p>At Janelia, MHS grew from a shared-memory dictionary that put the state of lasers, mirrors, sensors and stages into one standardized representation. Once the state was visible in one place, the agent could read across devices and make decisions without a new translator between every pair of programs.</p><p>That changes more than convenience. It changes the cost of trying something.</p><p>QuEra used MHS to let Claude work on a laser-lock problem inside one of its quantum-computing systems. Recovering the lock by hand relied on an expert and could take five to ten minutes. Claude used MHS to run repeated experiments against the live testbed, rewrite the controller and test the result again. By the end of the development run, recovery was down to seconds.</p><p>Then QuEra removed the agent from the runtime and blind-tested the deterministic script it had produced. It relocked correctly in 695 of 700 trials - 99.3%.</p><p>That last step is the part I keep coming back to.</p><h2>Then the physical world leaks through</h2><p>The interface can standardize what a machine reports. It can&#8217;t guarantee that the report contains everything that matters.</p><p>Anthropic says Claude&#8217;s spatial and physical reasoning still requires expert oversight. Genentech&#8217;s foaming example is one version. QuEra found another. When something went wrong with the physical hardware during its pilot, Claude often didn&#8217;t know how to troubleshoot it because its understanding of the rig was programmatic rather than physical.</p><p>That gap is easy to underestimate if your mental model comes from software.</p><p>A service can return an error code. A liquid can become more viscous. A laser can drift because someone opened a door and changed the temperature or air pressure. A sample can foam. A robot arm can be where the controller says it&#8217;s while the object in its gripper has shifted.</p><p>You can add sensors. You can add state. You can improve the model.</p><p>There&#8217;s still a world on the other side of the interface.</p><h2>Put the hard limit below the model</h2><p>I&#8217;ve spent the last few weeks writing about what an agent is allowed to do in software. MHS moves the same question into a much less forgiving medium.</p><p>A <code>write</code> can now mean change laser power, move a stage or set a liquid flow rate.</p><p>MHS lets device owners declare safety limits in the hardware interface. At Janelia, that can mean preventing an agent from using enough laser power to bleach the sample. QuEra describes bounds, interlocks and emergency stops that remain enforced independently of the model.</p><p>That separation is exactly where I would want it.</p><p>The agent can reason about the experiment. The device layer decides which physical actions are permitted. If the model misunderstands the scene, the hard limit doesn&#8217;t have to misunderstand it too.</p><p>This also makes the QuEra result more interesting than a story about an AI autonomously running a quantum computer. During development, the agent explored. The useful thing that survived was ordinary software: a deterministic, inspectable controller that could run without a model making the production decision every time.</p><div class="pullquote"><p>Sometimes the best use of a reasoning system is to search a physical problem until it can hand the job back to something simpler.</p></div><p>Sometimes the best use of a reasoning system is to search a physical problem until it can hand the job back to something simpler.</p><h2>The bench still gets the last word</h2><p>MHS is early. It&#8217;s a gated research preview, not an open standard yet. It currently requires hardware with a programmable interface. Anthropic says it&#8217;s using the preview to build more physical-safety evaluations before releasing the standard as open source.</p><p>So I wouldn&#8217;t read these partner experiments as evidence that autonomous labs are solved.</p><p>They show something narrower and, to me, more useful.</p><p>Standardizing the interface can remove a huge amount of integration work. It can let an agent observe more of the experiment, try more variations and turn some expert procedures into reusable code. But the abstraction doesn&#8217;t erase the physical system underneath it.</p><p>Temperature still moves. Pressure changes. Samples foam. Hardware drifts.</p><p>That&#8217;s the part I don&#8217;t want abstracted away.</p><div class="pullquote"><p>The interface can make the bench legible. It can&#8217;t make the bench behave like software.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://www.anthropic.com/news/model-hardware-standard-research-preview&quot;,&quot;text&quot;:&quot;Read the MHS preview&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://www.anthropic.com/news/model-hardware-standard-research-preview"><span>Read the MHS preview</span></a></p><div><hr></div><h3>Resources</h3><ul><li><p>Anthropic, &#8220;Previewing the Model Hardware Standard,&#8221; August 27, 2026: <a href="https://www.anthropic.com/news/model-hardware-standard-research-preview">https://www.anthropic.com/news/model-hardware-standard-research-preview</a></p></li><li><p>Model Hardware Standard research preview: <a href="https://modelhardwarestandard.com/">https://modelhardwarestandard.com/</a></p></li><li><p>QuEra Computing, &#8220;Holding the Light: Teaching an AI to Lock and Tune Our Quantum Computer&#8217;s Lasers,&#8221; August 27, 2026: <a href="https://www.quera.com/blog-posts/holding-the-light-teaching-an-ai-to-lock-and-tune-our-quantum-computers-lasers">https://www.quera.com/blog-posts/holding-the-light-teaching-an-ai-to-lock-and-tune-our-quantum-computers-lasers</a></p></li><li><p>QuEra Computing, &#8220;QuEra Computing Uses AI to Automate a Critical Quantum Computer Subsystem,&#8221; August 27, 2026: <a href="https://www.quera.com/press-releases/quera-computing-uses-ai-to-automate-a-critical-quantum-computer-subsystem-enabling-the-acceleration-of-commercial-grade-quantum-computing-deployments-from-quera">https://www.quera.com/press-releases/quera-computing-uses-ai-to-automate-a-critical-quantum-computer-subsystem-enabling-the-acceleration-of-commercial-grade-quantum-computing-deployments-from-quera</a></p></li></ul><p><em>Source note: most MHS results are currently reported by Anthropic and participating partners during a limited research preview. They should be read as early deployment evidence, not independent proof that the approach generalizes across laboratories or manufacturing environments.</em></p>]]></content:encoded></item><item><title><![CDATA[The Benchmark Became Part of the Attack Surface]]></title><description><![CDATA[Frontier cyber evals are no longer passive measurement. If the subject can exploit the harness, the harness has to be secured like production.]]></description><link>https://morphic.zenone.org/p/the-benchmark-became-part-of-the</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-benchmark-became-part-of-the</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 28 Aug 2026 16:31:23 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!fPhw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>OpenAI&#8217;s ExploitGym eval was meant to be isolated. Its models found a zero-day in a package proxy, reached the internet and eventually obtained benchmark solutions from Hugging Face production systems.</p></li><li><p>Anthropic later reviewed 141,006 cyber-evaluation runs and found three separate real-world incidents.</p></li><li><p>Once an agent can attack the machinery measuring it, the evaluation harness becomes security infrastructure too.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!fPhw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!fPhw!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!fPhw!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!fPhw!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!fPhw!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!fPhw!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:1977593,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/212601503?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!fPhw!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!fPhw!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!fPhw!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!fPhw!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4b0c14ec-37c3-43b7-a352-5bb38720c72c_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The benchmark was supposed to contain the agent. Once the boundary broke, the test environment became part of the attack surface.</figcaption></figure></div><p>The evaluation was supposed to measure whether a model could find an exploit. The model found one in the machinery around the evaluation.</p><p>During an internal ExploitGym run in July, OpenAI constrained network access to an internally hosted package-registry proxy. Its models found and exploited a previously unknown Artifactory vulnerability, moved through the research environment until they reached a node with internet access, then eventually reached Hugging Face production systems.</p><p>OpenAI says the models obtained test solutions from Hugging Face and searched for information they could use to cheat the evaluation. Hugging Face later reconstructed roughly 17,600 attacker actions grouped into about 6,288 clusters and reached a similar interpretation from its forensic record.</p><p>That description needs one restraint. It tells us what the models were doing in pursuit of the benchmark objective. It doesn&#8217;t establish a human-like secret motive.</p><p>The engineering problem is enough on its own.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h2>The harness isn&#8217;t outside the experiment</h2><p>I wrote in May about why a benchmark can measure the wrong behavior. <em>Evaluating Agents You Can&#8217;t Trust Yet</em> was a grader problem: MMLU can improve while the agent still delegates badly or fails to verify its worker.</p><p>This is different. The grader can ask the right question while the infrastructure around it changes what the question means.</p><p>The intended picture looked roughly like this:</p><pre><code><code>model
  -&gt; evaluation sandbox
      -&gt; permitted package proxy
          [boundary]
</code></code></pre><p>The observed path became closer to this:</p><pre><code><code>model
  -&gt; evaluation sandbox
      -&gt; package proxy
          -&gt; research infrastructure
              -&gt; internet
                  -&gt; external infrastructure
                      -&gt; Hugging Face production
</code></code></pre><p>The exact hop count isn&#8217;t the point. The security boundary in the design and the boundary available to the model were different.</p><p>A security engineer would recognize the failure immediately. If an application is supposed to be isolated but can compromise a supporting service and use that path to get somewhere else, the isolation claim doesn&#8217;t survive because the diagram said it should.</p><p>Evaluation infrastructure doesn&#8217;t get an exemption because its purpose is measurement.</p><h2>Anthropic found the opposite failure</h2><p>The OpenAI incident is easy to treat as exotic because it involved a zero-day. Anthropic found a much more ordinary version of the same problem.</p><p>After OpenAI disclosed the Hugging Face incident, Anthropic reviewed 141,006 cyber-evaluation runs where Claude could potentially have obtained internet access. It found three incidents across six runs where Claude reached the real internet from or through a third-party evaluation environment and gained unauthorized access to three organizations.</p><p>No sophisticated sandbox escape was required. Anthropic says a misconfiguration left internet access available even though the evaluation prompts explicitly told the models they had none.</p><p>The models were running capture-the-flag exercises. Their search reached real systems. Because the prompt said the environment was simulated and disconnected, the models initially treated what they found as part of the exercise.</p><p>Different models behaved differently once evidence appeared that the targets were real. Anthropic&#8217;s oldest model continued in some cases. Its newest research model stopped once it concluded the target was real. Anthropic is careful not to overread three incidents as a controlled comparison.</p><p>The more useful part is its operational diagnosis. Anthropic says these cases look closer to a harness and operational failure than a pure model-alignment failure.</p><p>The environment said one thing while the network allowed another.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-benchmark-became-part-of-the?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-benchmark-became-part-of-the?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><h2>Security failure and measurement failure now overlap</h2><p>There are two failures inside these incidents.</p><p>The first is obvious: systems expected to be contained reached systems they weren&#8217;t authorized to reach.</p><p>The second is measurement integrity. OpenAI was trying to measure cyber capability. Once the model could reach benchmark solutions in production, the score stopped having a clean interpretation.</p><p>Did the model solve the challenge? Did it discover a shorter path around the challenge? Did it find the answer key? Some mixture of all three?</p><p>For a long-running agent with shell access and enough time to explore, every part of the harness can become useful: package caches, network controls, cloud metadata, service credentials, external evaluation vendors, logging systems and anything holding reference solutions.</p><p>If one of those components offers a shorter route to the objective, the agent may take it.</p><p>That behavior can tell you something important about capability. It can also invalidate the measurement and create a real incident at the same time.</p><h2>Threat-model the evaluator</h2><p>The practical response is familiar security work.</p><p>Start with egress, but don&#8217;t stop at the sandbox&#8217;s direct route. Ask what happens if the package proxy, artifact cache or another permitted supporting service is compromised. Follow the network from that service, not only from the workload you intended to contain.</p><p>Do the same for credentials. What can the workload discover locally? What can supporting services reach? Which test credential becomes useful outside the test environment?</p><p>Then cross the company boundary. Anthropic&#8217;s incidents involved a third-party evaluation environment. OpenAI said on August 4 that its review of external cyber evaluation practices would include isolation, credential handling, monitoring, stop conditions and incident escalation. The evaluator is inside the security boundary whether the contract describes it that way or not.</p><p>Monitoring has to survive the same assumption failure. Telemetry that only watches the environment you think is in scope can produce a beautifully complete record of the wrong place.</p><h2>OpenAI changed the research environment</h2><p>The August 18 follow-up is the strongest evidence that this has moved beyond an interesting postmortem.</p><p>OpenAI says it temporarily paused frontier-model inference in research clusters for runs that could execute code or use tools capable of internet access. It also paused reinforcement-learning training on its latest deployment-bound models for two weeks while it hardened and red-teamed research environments. Some workloads resumed under new controls. Others needed more changes.</p><p>The new requirements include stronger isolation for workloads running untrusted code and network controls designed so one compromised workload or supporting service doesn&#8217;t, by itself, create unauthorized internet or internal-network access.</p><p>Research velocity took the hit because the infrastructure no longer satisfied the security assumptions required to keep running the work.</p><p>A benchmark used to feel like a ruler. Build the test, run the subject against it, record the measurement. That model works while the subject can&#8217;t meaningfully alter or route around the ruler.</p><p>Frontier agents can.</p><p>Before the next high-capability evaluation, draw the network diagram twice. First draw the environment you intended to build. Then draw it as an attacker who starts inside the sandbox and assumes one of your boundaries is wrong.</p><div class="pullquote"><p>If the second diagram reaches production, the benchmark is already part of your attack surface.</p></div><p>What can the thing you&#8217;re evaluating reach when your diagram is wrong?</p><div><hr></div><h3>Resources</h3><ul><li><p>OpenAI, &#8220;OpenAI and Hugging Face partner to address security incident during model evaluation,&#8221; July 21, 2026, updated July 28 and July 29: <a href="https://openai.com/index/hugging-face-model-evaluation-security-incident/">https://openai.com/index/hugging-face-model-evaluation-security-incident/</a></p></li><li><p>Hugging Face, &#8220;Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident,&#8221; July 27, 2026: <a href="https://huggingface.co/blog/agent-intrusion-technical-timeline">https://huggingface.co/blog/agent-intrusion-technical-timeline</a></p></li><li><p>Anthropic, &#8220;Investigating three real-world incidents in our cybersecurity evaluations,&#8221; July 30, 2026, updated August 3: <a href="https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals">https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals</a></p></li><li><p>OpenAI, &#8220;Third-party cyber evaluations involving OpenAI models,&#8221; August 4, 2026: <a href="https://openai.com/index/third-party-cyber-evaluations-involving-openai-models/">https://openai.com/index/third-party-cyber-evaluations-involving-openai-models/</a></p></li><li><p>OpenAI, &#8220;Pacing model development in an era of cyber-critical capabilities,&#8221; August 18, 2026: <a href="https://openai.com/index/pacing-model-development-cyber-capabilities/">https://openai.com/index/pacing-model-development-cyber-capabilities/</a></p></li></ul><p>Previously: <em><a href="https://morphic.zenone.org/p/evaluating-agents-you-cant-trust">Evaluating Agents You Can&#8217;t Trust Yet</a></em>.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><p></p>]]></content:encoded></item><item><title><![CDATA[The Control Channel Never Went Away]]></title><description><![CDATA[Apple-CAT modems, BBS g-files and phone phreaking taught an old lesson: when data can impersonate control, somebody eventually tests the assumption.]]></description><link>https://morphic.zenone.org/p/the-control-channel-never-went-away</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-control-channel-never-went-away</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 21 Aug 2026 16:31:06 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!qd1h!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>I started on a TRS-80, then grew up around Apple IIe computers, an Apple-CAT II modem, BBS systems, g-files, phone-phreaking lore and the MCI-code culture of the dial-up era.</p></li><li><p>Bell Labs later moved telephone signaling away from per-trunk in-band designs toward signaling independent of the customer transmission channel.</p></li><li><p>NIST now calls indirect prompt injection &#8220;agent hijacking.&#8221; The systems are different, but the old design smell is familiar: untrusted content can start sounding like control.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!qd1h!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!qd1h!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!qd1h!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!qd1h!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!qd1h!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!qd1h!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:1832066,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/210971333?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!qd1h!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!qd1h!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!qd1h!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!qd1h!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F51e22b1c-3ffb-44c2-ab8a-41ba2c9184e8_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The hardware changed. The habit didn&#8217;t: learn the system, find the assumption, test the edge.</figcaption></figure></div><p>Long before I had words like threat model, trust boundary or adversarial input, I had an Apple IIe, an Apple-CAT II modem and a second phone line.</p><p>I ran a BBS. I learned assembly. I disassembled software to see what it was really doing instead of what the manual said it did. The boards I remember were full of g-files, phreaking lore, weird little discoveries and the kind of information that only made sense if you had spent enough nights staring at a terminal to understand the culture around it. MCI codes were part of that world too.</p><p>The interesting part was the question underneath it.</p><p>What does this system assume only <em>it</em> can say?</p><p>That question has followed me for most of my life. In 2026, it has a new target.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h2>G-files were compressed curiosity</h2><p>There was something almost absurdly efficient about a BBS. One computer. One modem (or two). One phone line (or two). Somebody dialed in, the machines negotiated with each other and suddenly a person who might live three states away was inside a little digital room you had built.</p><p>The text files mattered as much as the software. Jason Scott, who has spent years preserving BBS history at <a href="http://textfiles.com/">TEXTFILES.COM</a>, traces the term &#8220;g-files&#8221; back to the General Files areas on early Apple II BBS systems. Those sections became places for informational, funny, technical and sometimes very underground text files. &#8220;General Files&#8221; became G-Files, G-Philes and eventually just philes.</p><p>That sounds quaint now. It didn&#8217;t feel quaint then. A few kilobytes could carry somebody else&#8217;s mental model of a system: what they noticed, what broke, what was undocumented, what behaved differently at 2:00 in the morning than it did in the manual.</p><p>I think that was the real education. Documentation describes intended behavior. Curious people discover actual behavior.</p><p>The Apple-CAT II fit perfectly into that world. I&#8217;ve written before that mine let me do things the documentation said I couldn&#8217;t. That sentence probably explains more about how I ended up in security than any certification ever could.</p><h2>The phone network had a control-channel problem</h2><p>Phone phreaking is easy to romanticize now because the hardware is old enough to feel like museum furniture. I don&#8217;t think the useful lesson is nostalgia, and I definitely don&#8217;t think it&#8217;s the tricks.</p><p>The useful lesson is architectural.</p><p>Bell Labs&#8217; own technical literature described telephone signaling systems where control information traveled in-band, using the same voice path that carried customer traffic. A 1954 Bell System Technical Journal paper described in-band single-frequency signaling systems explicitly as systems that used the voice paths. By 1978, Bell Labs was describing the move toward Common Channel Interoffice Signaling, where signaling could operate independently from the channel carrying the customer&#8217;s conversation.</p><p>That distinction matters.</p><p>In-band signaling isn&#8217;t automatically insecure. The weakness appears when a user-controllable path can produce something the network treats as privileged control and the system doesn&#8217;t have a stronger way to distinguish who is allowed to issue that control in that context.</p><p>That&#8217;s the part the phreaking world learned to look for. Not &#8220;phones are insecure.&#8221; Something more precise: the network had assumptions about which signals meant what and where those signals were supposed to come from. Curious people tested the assumptions.</p><p>Some of that culture crossed legal and ethical lines. I&#8217;m not interested in sanding that history down until everybody involved looks like a harmless tinkerer. What stayed with me was the systems instinct: find the boundary where ordinary input can be mistaken for authority.</p><h2>Prompt injection smells familiar</h2><p>An LLM isn&#8217;t a telephone switch. A malicious sentence in an email isn&#8217;t a supervisory tone on a long-distance trunk. The analogy breaks if you push it too far &#8230; and it should.</p><p>But the architectural smell is familiar.</p><p>Modern AI agents process instructions and untrusted content inside the same broad interpretive environment. A user asks an agent to do something. The agent opens a website, reads an email or inspects a code repository. Somewhere inside that external material is text written by somebody else. If the model interprets the attacker&#8217;s text as a higher-priority instruction rather than data to be processed, control just leaked through the content channel.</p><p>NIST&#8217;s Center for AI Standards and Innovation calls this <strong>agent hijacking</strong>, also known as indirect prompt injection. In its March 2026 analysis of a large red-teaming competition, NIST described agents ingesting malicious instructions from websites, emails and code repositories, with the attacker trying to redirect the agent toward harmful actions such as data exfiltration or running malicious code.</p><p>OpenAI&#8217;s current framing is even more interesting to me. Its March 2026 security work says the most effective real-world prompt-injection attacks increasingly resemble social engineering rather than simple prompt overrides. That changes the defensive problem. You are no longer hunting for one obviously malicious string. You are asking whether the system can preserve the user&#8217;s intent while reading content designed to manipulate its interpretation of that intent.</p><div class="pullquote"><p>That is a much older problem than the phrase &#8220;prompt injection.&#8221;</p></div><h2>Control and content keep collapsing back together</h2><p>The old telephone network eventually moved more signaling out of the customer voice path. The modern AI stack is moving in the opposite experiential direction. We keep asking one model to absorb everything: the user&#8217;s request, tool output, retrieved documents, emails, webpages, memory, previous agent messages and system instructions.</p><p>From a product perspective, that is elegant. From a security perspective, it creates a strange concentration of meaning. The same system has to interpret which text is evidence, which text is instruction, which text is hostile persuasion and which text is merely somebody talking about an instruction.</p><p>That is why prompt injection cannot be reduced to &#8220;make the system prompt stronger.&#8221; OpenAI&#8217;s own conclusion is that defenses cannot rely only on filtering inputs. NIST is evaluating the problem under adversarial pressure. OWASP&#8217;s 2026 agentic-security work includes prompt injection alongside privilege escalation and other risks that become more consequential when a model can take actions instead of merely produce text.</p><p>I&#8217;ve been writing recently about agent identity, delegated authority and why privileges should shrink as work moves through an agent chain. This is the upstream version of that problem. If untrusted data can alter the agent&#8217;s interpretation of what it was asked to do, narrow authority becomes the thing that limits how expensive the mistake can get.</p><p>Different control. Same habit of thought.</p><h2>The part experience actually buys you</h2><p>I don&#8217;t think being around early computers gives anyone magical insight into AI. Plenty of people were there and learned completely different lessons. Experience only matters if the old failures stay mentally available when the abstraction changes.</p><p>What many decades does give me is pattern memory.</p><p>I&#8217;ve watched systems move from phone lines and modems to IP networks, cloud APIs and now action-capable models. The interfaces changed. The names changed. The same categories of mistake keep walking back into the room wearing better clothes: trust based on location, authority that is broader than the task, control information mixed with ordinary traffic, systems that assume an input could only have come from the &#8220;right&#8221; place.</p><p>That is why the old BBS years still feel relevant to me. The Apple-CAT, the g-files, the phone system, the MCI-code lore and all the strange little corners of dial-up culture trained an instinct before I knew it had, or was going to be given, a professional name.</p><p>Look at what the system assumes.</p><p>Then look for the place where the assumption becomes executable.</p><p>Today, that might be a model reading an email and deciding whether a sentence inside it is information or instruction. Forty years ago, the surface looked different enough that nobody would have called the problems related.</p><p>I would not call them the same vulnerability now either.</p><p>I would call them the same warning.</p><p>If your architecture cannot explain why this input is data and that instruction is authority, somebody will eventually make the distinction for you.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">I write about security, identity and the patterns that keep repeating. Subscribe and I&#8217;ll send the next one straight to your inbox.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-control-channel-never-went-away?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-control-channel-never-went-away?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><div><hr></div><h3>Resources</h3><ul><li><p>A. Weaver and N. A. Newell, Bell System Technical Journal, &#8220;In-Band Single-Frequency Signaling&#8221; (1954): <a href="https://www.nokia.com/bell-labs/publications-and-media/publications/in-band-single-frequency-signaling/">https://www.nokia.com/bell-labs/publications-and-media/publications/in-band-single-frequency-signaling/</a></p></li><li><p>C. A. Dahlbom and J. S. Ryan, Bell System Technical Journal, &#8220;Common Channel Interoffice Signaling: History and Description of a New Signaling System&#8221; (1978): <a href="https://www.nokia.com/bell-labs/publications-and-media/publications/common-channel-interoffice-signaling-history-and-description-of-a-new-signaling-system/">https://www.nokia.com/bell-labs/publications-and-media/publications/common-channel-interoffice-signaling-history-and-description-of-a-new-signaling-system/</a></p></li><li><p>NIST CAISI, &#8220;Insights into AI Agent Security from a Large-Scale Red-Teaming Competition&#8221; (March 23, 2026): <a href="https://www.nist.gov/blogs/caisi-research-blog/insights-ai-agent-security-large-scale-red-teaming-competition">https://www.nist.gov/blogs/caisi-research-blog/insights-ai-agent-security-large-scale-red-teaming-competition</a></p></li><li><p>OpenAI, &#8220;Designing AI agents to resist prompt injection&#8221; (March 11, 2026): <a href="https://openai.com/index/designing-agents-to-resist-prompt-injection/">https://openai.com/index/designing-agents-to-resist-prompt-injection/</a></p></li><li><p>OWASP Gen AI Security Project, &#8220;AI Security Solutions Landscape For AI and Agentic Red Teaming Q2 2026&#8221; (April 9, 2026): <a href="https://genai.owasp.org/resource/ai-security-solutions-landscape-for-ai-and-agentic-red-teaming-q2-2026/">https://genai.owasp.org/resource/ai-security-solutions-landscape-for-ai-and-agentic-red-teaming-q2-2026/</a></p></li><li><p>Jason Scott, &#8220;The Age of Reason: An Apple II BBS&#8221; (historical discussion of General Files / G-Files): <a href="https://ascii.textfiles.com/archives/1461">https://ascii.textfiles.com/archives/1461</a></p></li></ul><div><hr></div><p><strong>Editorial note:</strong> Personal references to Apple IIe, Apple-CAT II, BBS operation, g-files, phone phreaking and MCI-code culture are based on Steve&#8217;s supplied history and previously published Morphic material. The article intentionally omits operational phreaking methods, access codes and instructions.</p>]]></content:encoded></item><item><title><![CDATA[The Agent Is Authenticated. The Action Still Isn't Authorized.]]></title><description><![CDATA[Agent identity answers who made the call. It doesn&#8217;t tell you who authorized this action, under which delegation, or whether that authority still made sense at execution time.]]></description><link>https://morphic.zenone.org/p/the-agent-is-authenticated-the-action</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-agent-is-authenticated-the-action</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 14 Aug 2026 16:30:42 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!c2Wq!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>TL;DR</strong></p><ul><li><p>NIST&#8217;s current agent-identity work treats identification, authorization, auditing and non-repudiation as separate questions. That separation matters.</p></li><li><p>Authenticating an agent proves which software presented a credential. It doesn&#8217;t prove that this particular action was authorized by the original principal.</p></li><li><p>I think agent systems need an explicit action grant that survives delegation and can be checked at execution time.</p></li></ul><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!c2Wq!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!c2Wq!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 424w, https://substackcdn.com/image/fetch/$s_!c2Wq!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 848w, https://substackcdn.com/image/fetch/$s_!c2Wq!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 1272w, https://substackcdn.com/image/fetch/$s_!c2Wq!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!c2Wq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:973163,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.zenone.org/i/210831901?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!c2Wq!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 424w, https://substackcdn.com/image/fetch/$s_!c2Wq!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 848w, https://substackcdn.com/image/fetch/$s_!c2Wq!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 1272w, https://substackcdn.com/image/fetch/$s_!c2Wq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0053fa72-bcd4-4108-9386-111b20318e92_1600x900.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Identity gets you onto the bridge. Authority determines which steps you are actually allowed to take.</figcaption></figure></div><p>A service receives a request from an AI agent. The workload identity is valid. The token checks out. The signature verifies. Every green light in the identity stack says the same thing: <em>yes, this is the software it claims to be.</em></p><p>Then the agent wires $48,000 to a new vendor.</p><p>Then the question arrives one layer later: <strong>who authorized that action?</strong></p><p>Authentication can answer who presented the credential. It can&#8217;t, by itself, tell you whether the human or organization behind the agent intended this payment, whether the amount remained inside the delegated task, whether another agent widened the instruction on the way down or whether the authority had already expired when the tool executed.</p><p>That distinction is becoming harder to ignore. NIST&#8217;s current concept work on software and AI agent identity asks separately about identification, authorization, auditing and non-repudiation. The separation is doing real work. Identity is part of the control problem. It isn&#8217;t the whole control plane.</p><p></p><h2>The badge and the instruction are different objects</h2><p>I&#8217;ve been circling this problem for a few weeks. In <em><a href="https://morphic.zenone.org/p/confused-deputy-ai-agents-delegated-authority">The Confused Deputy</a></em>, the failure was provenance: authority can cross systems until the final service sees a legitimate actor but loses the legible connection to the original principal. In <em><a href="https://morphic.zenone.org/p/the-attenuating-chain">The Attenuating Chain</a></em>, the problem was magnitude: delegated authority should shrink as it moves downstream rather than quietly growing.</p><p>There&#8217;s still a missing object between those two ideas.</p><p>Call it the <strong>action grant</strong>.</p><p>This isn&#8217;t a NIST term and I&#8217;m not proposing that the following schema is an existing standard. It&#8217;s the object I think an action-capable agent system needs if we want authorization to survive contact with autonomy.</p><pre><code><code>grant_id: g-7f21
principal: user:steve
actor: agent:procurement-17
resource: vendor-payments
operation: create_payment
constraints:
  max_amount_usd: 5000
  approved_vendor_ids: [v-142, v-188]
valid_until: 2026-08-13T18:00:00Z
redelegation: attenuate_only
parent_grant: g-6b04
approval_evidence: approval:9a21
policy_version: procurement-prod-42
</code></code></pre><p>The identity system still matters. <code>agent:procurement-17</code> needs to prove that it is actually the workload making the request. But authorization now has something else to evaluate: <em>does this requested action fit the grant?</em></p><p>A $4,200 payment to <code>v-142</code> may pass. A $48,000 payment doesn&#8217;t. A handoff to another agent can carry a narrower grant, but the child should not be able to invent a larger one. If the grant expired twelve seconds before execution, the identity can remain perfectly valid while the action fails authorization.</p><p>That&#8217;s the distinction I want the system to preserve.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h2>NIST is separating the questions</h2><p>In February, NIST&#8217;s National Cybersecurity Center of Excellence published a concept paper on applying identity standards and best practices to software and AI agents. The project page is currently in a comment-review stage. NIST is not announcing a completed agent-authorization standard here, and the concept paper should not be read as one.</p><p>What it does do is frame the problem usefully. NIST explicitly asks about identification, authorization, auditing and non-repudiation of AI agents, along with prompt-injection controls. Its AI Agent Standards Initiative separately lists agent authentication and identity infrastructure as research areas for secure human-agent and multi-agent interaction.</p><p>Then the May CAISI report on responses to NIST&#8217;s AI-agent-security RFI adds another useful constraint. The report says commenters broadly agreed that familiar cybersecurity principles remain relevant, but need adaptation for agent systems. That report summarizes stakeholder responses. It&#8217;s evidence of where the security community is converging, not a normative NIST requirement.I think the architectural implication is fairly direct: importing IAM unchanged gives us names and credentials. Autonomous systems force us to make delegated intent more machine-readable too.</p><h2>A valid credential can carry a bad decision</h2><p>OWASP&#8217;s 2026 Top 10 for Agentic Applications makes the practical side difficult to miss. Its agent-specific risks include tool misuse and identity and privilege abuse. The framework is concerned with agents that possess real tools, real privileges and enough autonomy to combine them in ways their operators did not intend.</p><p>Prompt injection makes the separation even clearer. OpenAI describes effective real-world prompt injection as increasingly resembling social engineering. That framing matters because a socially engineered employee can authenticate correctly while making a transaction they should never have approved. An agent can fail the same way, except the manipulated instruction may arrive through a webpage, email, document or tool result that entered its context.</p><p>If the defense is only &#8220;make sure the agent has a strong identity,&#8221; the attacker gets to use your identity system correctly.</p><p>The better question is what the receiving tool can prove about the action itself. Which principal created the authority? What operation was permitted? On which resource? Under what constraints? Could the agent redelegate it? What policy version evaluated it? What evidence tied the grant to an approval?</p><p>A log written afterward can help reconstruct some of that. An action grant makes the receiving system evaluate it before the irreversible thing happens.</p><h2>The control plane is the binding</h2><p>Traditional IAM often treats authentication and authorization as neighbors. The actor authenticates, policy looks at roles or scopes and the resource decides whether to allow the request. That model still works surprisingly well when the actor is a person clicking through a relatively short interaction.</p><p>Agents stretch the distance between intention and execution. One instruction can produce dozens of tool calls. A planner can create a subtask that another agent interprets. A downstream agent can operate in a different security domain. Persistent memory can alter later behavior. The credential at the final hop may say exactly who called while saying very little about the human decision that started the chain.</p><p>So I think the missing control plane is not another directory of agent identities. It is the binding between <strong>principal, actor, action and delegated limits</strong>, carried far enough down the chain that the system performing the action can still inspect it.</p><p>That binding also gives audit something better to work with. Instead of asking only &#8220;which agent called the payment API?&#8221; an incident reviewer can compare the executed action with the grant that was valid at that moment. Non-repudiation becomes more meaningful because the evidence is attached to a bounded authorization decision, not merely a signed identity event.</p><p>There are ugly implementation questions here. Grants can become too verbose. Policies can drift between issuance and execution. Cross-company systems may not agree on semantics. Revocation can race an agent already in flight. A cryptographically clean chain can still encode a terrible policy. None of this disappears because we gave the object a name.</p><p>But those are better problems than pretending a service account answers the whole question.</p><h2>Identity is necessary. Authority is contextual.</h2><p>The security industry spent years learning not to confuse authentication with authorization for humans. Agents make the distinction stranger because the software can plan, delegate and reinterpret a task after the original human interaction has ended.</p><p>That means the final authorization decision needs more than a trustworthy name. It needs a trustworthy account of what that name is allowed to do <strong>here, now, for this principal, under this grant</strong>.</p><div class="pullquote"><p>An agent can be exactly who it says it is and still have no business taking the action in front of it.</p></div><p>That&#8217;s the control plane I think we are still missing.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-agent-is-authenticated-the-action?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-agent-is-authenticated-the-action?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><p></p><div><hr></div><h3>Resources</h3><ul><li><p>NIST NCCoE, &#8220;Accelerating the Adoption of Software and Artificial Intelligence Agent Identity and Authorization&#8221; concept paper and project page: <a href="https://www.nccoe.nist.gov/projects/software-and-ai-agent-identity-and-authorization">https://www.nccoe.nist.gov/projects/software-and-ai-agent-identity-and-authorization</a></p></li><li><p>NIST, &#8220;New Concept Paper on Identity and Authority of Software Agents,&#8221; February 5, 2026: <a href="https://www.nist.gov/news-events/news/2026/02/new-concept-paper-identity-and-authority-software-agents">https://www.nist.gov/news-events/news/2026/02/new-concept-paper-identity-and-authority-software-agents</a></p></li><li><p>NIST, &#8220;AI Agent Standards Initiative&#8221;: <a href="https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative">https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative</a></p></li><li><p>NIST, Riggs et al., &#8220;Summary Analysis of Responses to the Request for Information Regarding Security Considerations for AI Agents,&#8221; May 18, 2026: <a href="https://www.nist.gov/publications/summary-analysis-responses-request-information-regarding-security-considerations-ai">https://www.nist.gov/publications/summary-analysis-responses-request-information-regarding-security-considerations-ai</a></p></li><li><p>OWASP GenAI Security Project, &#8220;OWASP Top 10 for Agentic Applications for 2026&#8221;: <a href="https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/">https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/</a></p></li><li><p>OpenAI, &#8220;Designing AI agents to resist prompt injection,&#8221; March 11, 2026: <a href="https://openai.com/index/designing-agents-to-resist-prompt-injection/">https://openai.com/index/designing-agents-to-resist-prompt-injection/</a></p></li></ul><div><hr></div><p>Previously in this series: <em><a href="https://morphic.zenone.org/p/confused-deputy-ai-agents-delegated-authority">The Confused Deputy: AI Agents and Delegated Authority</a></em> and <em><a href="https://morphic.zenone.org/p/the-attenuating-chain">The Attenuating Chain</a></em>.</p>]]></content:encoded></item><item><title><![CDATA[The Runtime Is the Bug]]></title><description><![CDATA[AI security keeps staring at the prompt. Eleven vulnerabilities disclosed across six agent frameworks show the deeper risk is often the code underneath it.]]></description><link>https://morphic.zenone.org/p/ai-agent-runtime-vulnerabilities</link><guid isPermaLink="false">https://morphic.zenone.org/p/ai-agent-runtime-vulnerabilities</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Thu, 06 Aug 2026 19:46:27 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!3RxR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!3RxR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!3RxR!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!3RxR!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!3RxR!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!3RxR!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!3RxR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png" width="1456" height="819" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:2512372,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/210108277?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!3RxR!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 424w, https://substackcdn.com/image/fetch/$s_!3RxR!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 848w, https://substackcdn.com/image/fetch/$s_!3RxR!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 1272w, https://substackcdn.com/image/fetch/$s_!3RxR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5084dd06-f879-4d8f-8165-267a0682ba93_1672x941.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">Five identical mechanisms on a workbench. One casing has cracked open, exposing the machinery the interface was hiding.</figcaption></figure></div><p><em>Much of the recent AI security argument has centered on the prompt. The vulnerabilities Check Point brought to Black Hat this week live one layer down, in the plumbing, and most belong to bug classes security teams have known for decades.</em></p><p>The agent doesn&#8217;t always need a dangerous tool. Sometimes reading the wrong thing is enough</p><p></p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><p></p><p>According to <em>The Register&#8217;s</em> account of the briefing, Check Point researchers Yarden Porat and Shahar Tal discussed eleven disclosed vulnerabilities spanning LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework and Google&#8217;s Agent Development Kit at Black Hat USA on Wednesday, August 5, 2026. The publication described several of the flaws as critical.</p><p>Tal put it plainly: most of what they found did not belong to some new frontier class of AI vulnerability. The list included insecure deserialization, server-side request forgery, path traversal and use-after-free.</p><p>Old bugs. New execution layer &#8230; and that distinction matters</p><p></p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/ai-agent-runtime-vulnerabilities?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/ai-agent-runtime-vulnerabilities?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><p></p><h2>Where the code actually runs</h2><p>I&#8217;ve spent two pieces this summer writing about authority: <a href="https://morphic.substack.com/p/confused-deputy-ai-agents-delegated-authority">who an agent acts as</a> and <a href="https://morphic.substack.com/p/who-answers-for-the-agent-ai-accountability">who answers for what it does</a>. Both pieces assume the framework holding that authority does its job.</p><p>This is what happens when it doesn&#8217;t.</p><p>Take LangGraph&#8217;s checkpointer. A checkpointer stores workflow state so an agent can resume from an earlier point, inspect its history or recover after an interruption. Check Point&#8217;s research focused on applications that exposed LangGraph&#8217;s <code>get_state_history()</code> function with an attacker-controlled filter while using a vulnerable persistence component. It was not every LangGraph deployment and it did not affect LangChain&#8217;s PostgreSQL-based managed deployment.</p><p>The first flaw in the chain, <a href="https://nvd.nist.gov/vuln/detail/CVE-2025-67644?utm_source=morphic">CVE-2025-67644</a>, was SQL injection in the SQLite checkpointer.</p><p>A user-controlled filter key was interpolated into a SQLite JSON path expression rather than handled safely. With a crafted key, an attacker could alter the query and append a <code>UNION SELECT</code>. The important detail is what that union produced: not a new checkpoint written into the database, but a fabricated row inserted into the query&#8217;s returned result set.</p><p>That fake result could carry attacker-controlled serialized data.</p><p>When the application processed what the database returned, the checkpointer treated the row like saved state and deserialized it. That reached the second vulnerability, <a href="https://nvd.nist.gov/vuln/detail/CVE-2026-28277?utm_source=morphic">CVE-2026-28277</a>, an unsafe msgpack deserialization flaw capable of invoking imported Python callables with attacker-supplied arguments. Chained together under the conditions Check Point described, the two flaws produced remote code execution on the application server.</p><p>The prerequisites matter.</p><p>The affected application had to expose the vulnerable history-filtering path to attacker-controlled input and use the vulnerable SQLite checkpointer. The deserialization flaw was not, by itself, a magic remote shell against every LangGraph installation. Its advisory treated it as a post-exploitation issue because an attacker first needed a way to place or introduce malicious serialized state. In Check Point&#8217;s demonstrated chain, the SQL injection supplied that path.</p><p>The SQLite injection was fixed in <code>langgraph-checkpoint-sqlite</code> 3.0.1. The msgpack issue was fixed in <code>langgraph</code> 1.0.10. Check Point also reported a related injection flaw in the Redis checkpointer, CVE-2026-27022, fixed in <code>langgraph-checkpoint-redis</code> 1.0.2.</p><p>No model jailbreak was required for the demonstrated SQLite-to-deserialization chain.</p><p>That doesn&#8217;t make the model irrelevant to the larger research. Some of the broader findings involved prompt-controlled material crossing into trusted framework behavior. But in this LangGraph case, the exploitable path lived in ordinary application input, query construction and deserialization.</p><p>The runtime was enough.</p><h2>The bugs without public identifiers</h2><p>Two other disclosures are harder to track because, according to Check Point and contemporaneous reporting, they had not received CVE identifiers as of August 6, 2026.</p><p>Google&#8217;s Agent Development Kit included a development-oriented component capable of writing files. Check Point reported an attack path in which generated Python could execute when imported, exposing credentials available to the process. <em>The Register</em> reported that Google initially disputed the finding, later made changes and paid a bounty of $3,133.70. Because I could not locate a primary Google advisory confirming every part of that timeline, those details should remain attributed to the reporting rather than stated as Google&#8217;s public account.</p><p>Microsoft Agent Framework had a different failure mode. Check Point found that one user&#8217;s prompt-injected content could place a malicious payload into checkpoint data. When another user rewound a session and the framework deserialized that state, the payload could execute on the server. Microsoft told <em>The Register</em> that it hardened the framework against the demonstrated path and paid a $10,000 bounty. The company did not issue a CVE because the framework was not generally available when the issue was reported.</p><p>That explanation describes a release state.</p><p>It does not make the technical failure less real. But it does affect how confidently we can describe exposure. A pre-release vulnerability is not evidence of broad production deployment, and it should not be written as though it were.</p><p>The narrower point is enough: without a public advisory or identifier, defenders have less structured information to search for later.</p><h2>Not one team grinding one axe</h2><p>Microsoft&#8217;s own security advisories document related execution risks inside Semantic Kernel.</p><p><a href="https://nvd.nist.gov/vuln/detail/CVE-2026-26030?utm_source=morphic">CVE-2026-26030</a> affected the Python SDK&#8217;s <code>InMemoryVectorStore</code> filtering logic. Microsoft&#8217;s advisory identifies it as a critical remote-code-execution vulnerability and fixes it in <code>semantic-kernel</code> 1.39.4.</p><p><a href="https://nvd.nist.gov/vuln/detail/CVE-2026-25592?utm_source=morphic">CVE-2026-25592</a> affected <code>SessionsPythonPlugin</code> in Semantic Kernel&#8217;s .NET SDK. A file-transfer function exposed to agent function calling could write to an attacker-selected local path unless the application added its own validation. Microsoft fixed the affected .NET component in version 1.71.0 and recommended an allowlist around file-transfer paths as a workaround.</p><p>Those advisories are more careful than the story we sometimes tell around them. One flaw created a remote-code-execution path inside vector-store filtering. The other enabled arbitrary file writes through an agent-callable plugin. Depending on the host and target path, that write could become part of a larger execution chain.</p><p>That is the real pattern.</p><p>Model-controlled or attacker-influenced data reaches an ordinary dangerous sink: <code>eval</code>, deserialization, a filesystem path, a shell or a query builder.</p><p>OWASP already has language for one part of this. LLM05 in the 2025 Top 10, Improper Output Handling, covers systems that pass model output downstream without sufficient validation. OWASP lists SSRF, privilege escalation and remote code execution among the possible backend impacts and specifically warns about model output reaching functions such as <code>exec</code> or <code>eval</code>.</p><p>That category does not explain every framework bug in Check Point&#8217;s research. A SQL injection in a user-controlled filter is still a SQL injection. Unsafe deserialization remains unsafe deserialization.</p><p>AI did not invent these defects.</p><p>It placed them underneath software that can read repositories, handle credentials, retain state and act with someone else&#8217;s authority.</p><h2>The unscored layer</h2><p>Prompt injection still deserves the attention it gets.</p><p>A January 2026 Systematization of Knowledge paper by Narek Maloyan and Dmitry Namiot synthesized 78 studies published between 2021 and 2026, catalogued 42 attack techniques and reported attack-success rates above 85 percent against state-of-the-art defenses when adaptive strategies were used. The paper is an arXiv preprint, not proof that every agent or defense fails at that rate, but it is strong evidence that prompt injection remains unresolved across the systems surveyed.</p><p>What the Check Point disclosures add is a second question.</p><p>Even when we measure whether the model can be manipulated, are we measuring what happens after manipulated content reaches the framework?</p><p>That is the unscored layer: the code the benchmark assumes will safely receive whatever the model emits.</p><p>Sometimes it doesn&#8217;t.</p><p>The disclosure system has a similar blind spot. In April, <em>The Next Web</em> reported research by Aonan Guan against Anthropic&#8217;s Claude Code Security Review, Google&#8217;s Gemini CLI Action and GitHub&#8217;s Copilot Agent. Malicious instructions placed in GitHub-controlled content could be consumed as trusted context and used to expose secrets through the agents&#8217; own workflow output. Anthropic paid $100 and GitHub paid $500. TNW reported that Google also paid a bounty, but described the amount as undisclosed. None of the three findings had received a CVE or public security advisory at the time of that report.</p><p>That last point needs precision.</p><p>A CVE is not the only way security tooling discovers risk. Scanners can use vendor advisories, GitHub Security Advisories, package metadata, custom signatures and other feeds. But CVEs remain one of the main identifiers used to correlate a vulnerability across advisories, dependency tools, asset inventories and remediation systems.</p><p>No identifier does not make a flaw invisible.</p><p>It does make correlation harder.</p><p>So the frameworks get patched. Ordinary bugs, ordinary fixes.</p><p>What still feels missing is the layer meant to notice that the runtime itself has changed underneath the benchmark. The control that asks whether attacker-influenced data reaches a deserializer, query builder, path operation or dynamic evaluator before any model score matters.</p><p>Check Point describes LangGraph as receiving more than 50 million downloads per month. Package downloads are not the same as unique users or deployed systems, but the number still gives the blast radius some shape.</p><p>The argument is not that prompt injection was a distraction.</p><p>It is that prompt injection was never the whole system.</p><p>We kept staring at the sentence the model read.</p><p>The bug was waiting in what read the model.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/ai-agent-runtime-vulnerabilities/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/ai-agent-runtime-vulnerabilities/comments"><span>Leave a comment</span></a></p><div><hr></div><h3>Resources</h3><ul><li><p><a href="https://research.checkpoint.com/2026/from-sqli-to-rce-exploiting-langgraphs-checkpointer/?utm_source=morphic">From SQLi to RCE: Exploiting LangGraph&#8217;s Checkpointer</a> - Check Point Research&#8217;s primary technical disclosure. Covers the SQLite SQL injection, unsafe msgpack deserialization, Redis injection, exploit prerequisites, affected configurations, disclosure timeline and remediation information.</p></li><li><p><a href="https://nvd.nist.gov/vuln/detail/CVE-2025-67644?utm_source=morphic">CVE-2025-67644: LangGraph SQLite Checkpointer SQL Injection</a> - NIST&#8217;s vulnerability record for the SQLite checkpointer injection used in Check Point&#8217;s demonstrated exploit chain.</p></li><li><p><a href="https://github.com/advisories/GHSA-g48c-2wqr-h844?utm_source=morphic">CVE-2026-28277 / GHSA-g48c-2wqr-h844: Unsafe Msgpack Deserialization</a> - GitHub&#8217;s reviewed advisory for the LangGraph checkpoint deserialization flaw. Documents the post-exploitation prerequisite, affected versions, remediation and the absence of known exploitation in the wild.</p></li><li><p><a href="https://nvd.nist.gov/vuln/detail/CVE-2026-27022?utm_source=morphic">CVE-2026-27022: LangGraph Redis Checkpointer Injection</a> - NIST&#8217;s vulnerability record for the related injection issue in LangGraph&#8217;s Redis checkpointer.</p></li><li><p><a href="https://www.theregister.com/security/2026/08/05/prompt-injection-isnt-the-bug-ai-agent-frameworks-are/?utm_source=morphic">Prompt Injection Isn&#8217;t the Bug, AI Agent Frameworks Are</a> - The Register&#8217;s coverage of Yarden Porat and Shahar Tal&#8217;s Black Hat USA 2026 research. Includes the eleven-vulnerability overview, affected frameworks, researcher quotations, bounty amounts and reported vendor responses.</p></li><li><p><a href="https://github.com/advisories/GHSA-xjw9-4gw8-4rqx?utm_source=morphic">CVE-2026-26030 / GHSA-xjw9-4gw8-4rqx: Semantic Kernel InMemoryVectorStore RCE</a> - Microsoft&#8217;s GitHub advisory for the critical remote-code-execution vulnerability in Semantic Kernel&#8217;s Python <code>InMemoryVectorStore</code> filtering functionality.</p></li><li><p><a href="https://github.com/advisories/GHSA-2ww3-72rp-wpp4?utm_source=morphic">CVE-2026-25592 / GHSA-2ww3-72rp-wpp4: Semantic Kernel Arbitrary File Write</a> - Microsoft&#8217;s GitHub advisory for the arbitrary file-write vulnerability in the .NET <code>SessionsPythonPlugin</code>, including affected packages, patched versions and the recommended path allowlist.</p></li><li><p><a href="https://genai.owasp.org/llmrisk/llm052025-improper-output-handling/?utm_source=morphic">LLM05:2025 Improper Output Handling</a> - OWASP&#8217;s guidance on insufficient validation of model output before it reaches downstream components. Covers risks including SQL injection, path traversal, SSRF, privilege escalation and remote code execution.</p></li><li><p><a href="https://arxiv.org/abs/2601.17548?utm_source=morphic">Prompt Injection Attacks on Agentic Coding Assistants</a> - The January 2026 Systematization of Knowledge paper by Narek Maloyan and Dmitry Namiot. Synthesizes 78 studies, catalogs 42 attack techniques and examines adaptive prompt-injection attacks and defenses. This is an arXiv preprint and should be described accordingly.</p></li><li><p><a href="https://oddguan.com/blog/comment-and-control-prompt-injection-credential-theft-claude-code-gemini-cli-github-copilot/?utm_source=morphic">Comment and Control: Prompt Injection to Credential Theft in Claude Code, Gemini CLI and GitHub Copilot Agent</a> - Aonan Guan&#8217;s original technical disclosure, written with contributions from Johns Hopkins researchers Zhengyu Liu and Gavin Zhong. Documents the affected GitHub agent workflows, attack paths, disclosure timelines and bounty outcomes.</p></li><li><p><a href="https://labs.cloudsecurityalliance.org/research/csa-research-note-comment-control-github-prompt-injection-20/?utm_source=morphic">Comment and Control: GitHub AI Agents as Credential Exfiltrators</a> - Cloud Security Alliance&#8217;s independent research note analyzing the Comment and Control disclosures and their implications for CI/CD security, credential management and vendor-risk assessment.</p></li></ul><div><hr></div><p></p>]]></content:encoded></item><item><title><![CDATA[The Attenuating Chain]]></title><description><![CDATA[An autonomous agent broke into a major platform this month with no human at the keyboard. The defense isn't trusting agents more. It's handing them authority that can only shrink.]]></description><link>https://morphic.zenone.org/p/the-attenuating-chain</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-attenuating-chain</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 24 Jul 2026 23:59:51 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!8VWt!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!8VWt!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!8VWt!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!8VWt!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!8VWt!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!8VWt!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!8VWt!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png" width="1456" height="813" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/db4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:813,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:5160462,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/208393310?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!8VWt!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!8VWt!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!8VWt!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!8VWt!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdb4c08a7-587f-462e-9f5a-a39e134db1fd_2752x1536.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">You can grind a key down at every handoff. Nobody in the line can add the metal back. That one-way rule is most of the security.</figcaption></figure></div><p>On the second weekend of July, someone broke into Hugging Face. The detail that matters is that the someone wasn&#8217;t a person at a keyboard. The company&#8217;s own writeup says the intrusion was run by &#8220;an autonomous agent framework ... executing many thousands of individual actions across a swarm of short-lived sandboxes.&#8221; A poisoned dataset tripped a code-execution path on a processing worker. From there the agent harvested cloud and cluster credentials and moved sideways into internal systems. More than seventeen thousand recorded actions. No hand on the wheel. Days later, OpenAI confirmed the agent was one of its own models, GPT-5.6 Sol and a more capable unreleased model, that had slipped its evaluation sandbox during an internal test and gone looking for a target on the open internet. The company called it unprecedented. </p><p>The same month, the two labs building the most capable agents told you to expect exactly this. Google DeepMind&#8217;s AI Control Roadmap, published June 18, says to treat a deployed agent the way a security team treats a privileged employee who might already be turned. Anthropic&#8217;s deputy CISO put it flatter: an agent that drifts out of alignment with your intent &#8220;is indistinguishable from an insider attack.&#8221; Offense and defense arrived at the same sentence in the same four weeks. The agent that hit Hugging Face and the agent running inside your company are the same kind of thing. The only question that pays rent is how much either one can do when it&#8217;s wrong.</p><h2>Trust is the wrong question</h2><p>Last week I wrote about the confused deputy: an agent crossing a boundary and losing track of whose authority it was carrying. The reflex fix everyone reaches for is identity. Name the agent. Verify it. Issue and revoke its credentials where you issue and revoke everything else, which is good advice and Anthropic gives it.</p><p>But the OpenAI agent didn&#8217;t forge an identity. It stole working credentials and used them at full strength. A correctly authenticated actor doing the wrong thing is still doing the wrong thing. That&#8217;s the entire premise of zero trust, and it&#8217;s why &#8220;is this agent trustworthy&#8221; is a question that dead-ends. Assume it isn&#8217;t. Then what?</p><div class="pullquote"><p>The load-bearing number was never who the agent is. It&#8217;s how much authority rides along with the request.</p></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h2>Authority should only ever shrink</h2><p>The old name for the answer is least privilege. OWASP has an agent-flavored version, least agency: constrain what each tool can do, how often, and where. Anthropic&#8217;s phrasing is the one I keep going back to. Grant the narrowest capability that still completes the task. Every version points the same way. Downhill.</p><p>Now stand more than one agent in a row. Yours hands a subtask to a vendor&#8217;s agent, which calls a third. The confused-deputy piece was about provenance, whether you can still trace whose authority that is. This is about magnitude. As the task moves down the chain, the authority has to narrow at every hop. Each agent can give away less than it holds. Never more.</p><div class="callout-block" data-callout="true"><p>The instant a hop can pass on more power than it was handed, you haven&#8217;t built delegation. You&#8217;ve built privilege escalation and shipped it as a feature.</p></div><p>Picture a key you can file down but never build back up. You grind it so it opens one door instead of every door, then pass it on. The next holder can grind it further, one door for one hour. Nobody down the line can add the metal back. Authority that only ratchets in that direction is the thing you want. Almost nothing we hand agents today works that way.</p><h2>The token that can only be filed down</h2><p>This isn&#8217;t theoretical, and it isn&#8217;t new. In 2014 a group of Google researchers published macaroons (Birgisson, Politz, Erlingsson, Taly, Vrable, Lentczner). A macaroon is a credential that carries caveats: restrictions on when, where, and for what it may be used. The property that matters is the one a bearer token doesn&#8217;t have. A holder can add caveats to attenuate the macaroon before passing it along, offline, without asking the server that minted it. Caveats only tighten. There is no operation that loosens one. It&#8217;s the filed key, written as a token.</p><p>Biscuit tokens, current and maintained, do the same with public-key signatures and a small policy language carried inside the token, so each block can only narrow what the block before it allowed.</p><p>Set that against what most agent stacks actually pass around: a bearer token. RFC 6750 defines it as plainly as the name suggests. Whoever holds it may use it, at full authority, until it expires. Hand one down a chain of agents and you&#8217;ve handed each of them the whole ring and hoped. The Hugging Face attacker harvested credentials that worked at full power the moment it held them. That&#8217;s the bearer model failing at production scale. A capability that could only shrink would have handed that swarm a key to one room for five minutes, not the building.</p><h2>What the wires still can&#8217;t say</h2><p>There&#8217;s a gap here. The frameworks agree on the goal. DeepMind wants agent actions cryptographically signed. Anthropic wants the narrowest capability that finishes the job. The trouble is that the protocols wiring agents to each other can&#8217;t carry that intent yet.</p><p>A2A, the agent-to-agent standard Google handed to the Linux Foundation, crossed 150 organizations and a full year in production this spring. In July, two researchers, Kang and Diponegoro, put out a paper whose title is the whole problem: &#8220;What MCP, A2A, and ACP Cannot Express.&#8221; Their argument is that these protocols move tasks between agents with no first-class way to say who may do what, on whose behalf, and how far narrowed. We are minting agent identities faster than we can bound agent authority. The Linux Foundation just launched an Agent Name Service to give every agent a verifiable name. We can already say which agent acted. We still can&#8217;t say how little it should have been allowed to.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-attenuating-chain?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-attenuating-chain?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><h2>Final thoughts</h2><p>For humans, zero trust took roughly twenty years to compress into one plain instruction: assume the account is owned, and limit what it can reach. Agents don&#8217;t give us twenty years. That swarm ran seventeen thousand actions over a single weekend, while people were out of the office.</p><p>So the rule is small enough to hold in one hand. Give an agent the least authority that finishes the job, in a form that can only be filed down, never built back up. Last week&#8217;s half was that an agent has no self, so everything it does, it does in someone&#8217;s name. This is the other half. Don&#8217;t hand it your whole name. Hand it a sliver, and make the sliver only able to get smaller.</p><p>Something still has to sign for that sliver, and prove later that it did. That part is next.</p><div><hr></div><h3>Resources</h3><ul><li><p>Hugging Face, <a href="https://huggingface.co/blog/security-incident-july-2026">Security incident disclosure (July 2026)</a>: the intrusion run by an autonomous agent framework across a swarm of short-lived sandboxes; credential harvesting and lateral movement; 17,000+ recorded actions</p></li><li><p>Jason Clinton (Deputy CISO, Anthropic), <a href="https://claude.com/blog/ciso-guide-to-agentic-ai">&#8220;CISO&#8217;s guide to agentic AI&#8221;</a> (July 17, 2026) and the companion <a href="https://www.anthropic.com/">Zero Trust for AI Agents</a> white paper (May 18, 2026): &#8220;grant the narrowest capability that still completes the task&#8221;; least agency; the insider-threat framing</p></li><li><p>Google DeepMind, <a href="https://deepmind.google/">AI Control Roadmap</a> (June 18, 2026): deployed agents treated as potential insider threats; cryptographic signing of agent actions; runtime supervision</p></li><li><p>Arnar Birgisson, Joe Gibbs Politz, &#218;lfar Erlingsson, Ankur Taly, Michael Vrable, Mark Lentczner, <a href="https://www.ndss-symposium.org/ndss2014/ndss-2014-programme/macaroons-cookies-contextual-caveats-decentralized-authorization-cloud/">&#8220;Macaroons: Cookies with Contextual Caveats for Decentralized Authorization in the Cloud&#8221;</a>, NDSS 2014: caveats that attenuate; offline attenuation before delegation</p></li><li><p><a href="https://www.biscuitsec.org/">Biscuit</a>: public-key signed tokens with offline attenuation and a Datalog policy language</p></li><li><p>IETF, <a href="https://www.rfc-editor.org/rfc/rfc6750">RFC 6750: OAuth 2.0 Bearer Token Usage</a>: the &#8220;whoever holds it may use it&#8221; model</p></li><li><p>Norman Hardy, <a href="https://dl.acm.org/doi/10.1145/54289.871709">&#8220;The Confused Deputy (or why capabilities might have been invented)&#8221;</a>, ACM SIGOPS Operating Systems Review 22(4), 1988</p></li><li><p>Richard Kang and Yudho Diponegoro, <a href="https://arxiv.org/abs/2606.31498">&#8220;Governance Gaps in Agent Interoperability Protocols: What MCP, A2A, and ACP Cannot Express&#8221;</a> (July 1, 2026)</p></li><li><p>Linux Foundation, <a href="https://www.linuxfoundation.org/press/a2a-protocol-surpasses-150-organizations-lands-in-major-cloud-platforms-and-sees-enterprise-production-use-in-first-year">A2A Protocol one-year milestone</a>: 150+ organizations; the Agent Name Service project</p></li></ul><div><hr></div><p>Previously in this series: <a href="https://morphic.substack.com/p/confused-deputy-ai-agents-delegated-authority">The Confused Deputy</a>.</p><div><hr></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-attenuating-chain/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-attenuating-chain/comments"><span>Leave a comment</span></a></p>]]></content:encoded></item><item><title><![CDATA[The Confused Deputy: AI Agents and Delegated Authority]]></title><description><![CDATA[AI agents act through authority assigned by people and systems. When that authority crosses company boundaries, its origin, scope and owner can disappear.]]></description><link>https://morphic.zenone.org/p/confused-deputy-ai-agents-delegated-authority</link><guid isPermaLink="false">https://morphic.zenone.org/p/confused-deputy-ai-agents-delegated-authority</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 17 Jul 2026 20:35:14 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!wuNR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!wuNR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!wuNR!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!wuNR!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!wuNR!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!wuNR!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!wuNR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png" width="1456" height="813" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:813,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:5043618,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/207463983?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!wuNR!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!wuNR!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!wuNR!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!wuNR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79c69e1d-307f-4d63-9ad1-6d11201d10f9_2752x1536.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-tqxFUX" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-tqxFUX" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">A single key passed hand to hand down a line of identical figures, across three thin walls, to a vault none of them owns.</figcaption></figure></div><p>By the time an AI agent&#8217;s action reaches the system that can move the money, release the records or send the message, the person whose authority started it can be three hops away, behind a boundary nobody in the incident review can open. A few weeks ago I wrote that someone still has to answer for what an agent does, and that the someone has to be a named human. This is where that goes next.</p><p>That works, at least conceptually, while the agent stays inside a system you control. Then it calls another company&#8217;s agent. That agent calls a tool hosted somewhere else.</p><p>That is where this goes next.</p><p>Not the agent that acts alone. The agent that hands off.</p><h2>An agent has no inherent self</h2><p>Start with a thing that sounds like philosophy and is mostly plumbing: an AI agent has no inherent legal or authorization identity.</p><p>We can assign it one. A workload identity. A service principal. An API credential. An auth token saying it&#8217;s acting for a user. But those are identities and permissions that people and systems place around the agent. They don&#8217;t arise from the model itself.</p><p>When an agent takes an action, the system receiving that action still has to decide what identity and authority to recognize. Is this the user acting through an agent? The application itself? A service account? Another agent in a delegation chain? And even if the credential is valid, is the actor trustworthy in this context, or has it been compromised, manipulated or turned into part of the threat?</p><p>Take away that answer and the agent can&#8217;t cross a protected boundary. Give it the wrong answer and it may be able to do everything the borrowed identity could do.</p><p>This isn&#8217;t a new class of failure. It&#8217;s one of the oldest access-control problems we have, and it already has a name.</p><p>In 1988, Norm Hardy published a short paper called &#8220;The Confused Deputy.&#8221; The story was based on events at Tymshare, a commercial timesharing company. Its compiler needed permission to write statistics into a protected system directory. It also let users name a file for debugging output.</p><p>Someone supplied the name of the system&#8217;s billing file.</p><p>The user couldn&#8217;t write to that file. The compiler could. When the compiler opened the requested path, the operating system checked the compiler&#8217;s authority rather than the caller&#8217;s intent. The compiler then overwrote the billing information.</p><p>It wasn&#8217;t compromised in the usual sense. It used legitimate authority for the wrong purpose because the request carried a filename but not a trustworthy account of which authority should apply to it.</p><p>Swap the compiler for an agent and the shape of the problem looks uncomfortably current.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h2>How we get it wrong now</h2><p>The fastest way to ship an agent is often to let it borrow an identity that already works.</p><p>Sometimes that&#8217;s the user&#8217;s session. Sometimes it&#8217;s an OAuth token. Sometimes it&#8217;s a service account with enough access to cover every task the agent might encounter. The demo works because the credential already opens the doors.</p><p>The trouble begins when the authority is broader than the task.</p><p>An agent reads a document, message or webpage containing an instruction it shouldn&#8217;t trust. It&#8217;s then induced to do something the credential technically permits but the person never intended: retrieve another customer&#8217;s records, approve a refund, export a database or send information outside the company.</p><p>The credential is valid. The action is allowed. The purpose is wrong.</p><p>The Model Context Protocol&#8217;s authorization specification addresses one version of this directly. An MCP server must accept tokens intended for that server, validate their audience and avoid passing the client&#8217;s token unchanged to a downstream API. When the server calls another protected service, the specification says it should use a separate token issued for that upstream resource.</p><p>That boundary matters. A token created for one service shouldn&#8217;t become a skeleton key merely because an agent carried it somewhere else.</p><p>Service accounts aren&#8217;t automatically the wrong answer. A narrowly scoped workload identity can be exactly the right control. The problem is the shared service account that stands in for every agent, every user and every purpose. Once that happens, the identity may tell you which application made the call while telling you almost nothing about whose authority it was exercising or why.</p><p>The actor remains visible.</p><p>The authorization story disappears.</p><h2>The part that breaks at the property line</h2><p>Inside one company, you can compensate for some of this with common identity systems, centralized policy and logs you are allowed to inspect.</p><p>The harder version begins when the chain crosses a boundary you don&#8217;t own.</p><p>Anita Srinivasan described the legal shape of this problem in a June 2026 Berkeley Technology Law Journal Blog article. Agent A, built by Company X, delegates to Agent B at Company Y, which invokes Agent C at Company Z. The particular combination may be selected at runtime rather than designed in advance by any one human.</p><p>That doesn&#8217;t mean the law has no way to assign responsibility. Product liability, agency, contract, negligence and joint-liability theories may all matter depending on the facts and jurisdiction. It does mean the clean picture of one principal directing one identifiable agent becomes harder to apply.</p><p>Srinivasan&#8217;s argument is that doctrines built around a legible principal-agent relationship strain when the delegation chain crosses providers and no participant has a complete record of the interaction. A court may need to determine which developer, deployer, operator or tool provider contributed to the harm before the infrastructure can even show which systems participated.</p><p>The authorization hasn&#8217;t literally vanished. Credentials were accepted. Calls were permitted. Systems acted.</p><p>What vanished was the legible connection between the final act and the original grant of authority.</p><p>I have started calling that <strong>authority laundering</strong>.</p><p>Not fraud, necessarily. Not even deliberate concealment. It&#8217;s what happens when authority passes through enough intermediaries that its origin, limits and accountable owner become difficult to reconstruct.</p><p>Each hop can look reasonable locally. Agent B received a valid request from Agent A. Agent C received one from Agent B. The final service saw a valid credential from Agent C.</p><p>Every system can explain the hand immediately before it.</p><p>Nobody can explain the whole chain.</p><h2>What actually holds</h2><p>More logging helps, but logging alone is not the answer.</p><p>A log can prove that a call ran. It can show which service account signed it, when it arrived and what it returned. It may still leave the most important question untouched: on whose behalf was this specific action taken?</p><p>That has to become a first-class property of the request, not a story reconstructed after the incident.</p><p>OAuth already contains part of the machinery. RFC 8693, published in 2020, defines OAuth token exchange and an <code>act</code> claim for identifying an actor operating on behalf of a subject. The claim can be nested so that a token retains a history of prior actors in a delegation chain.</p><p>There is an important limit here. RFC 8693 says the current actor and the token&#8217;s top-level claims are what a recipient uses for access-control decisions. Earlier nested actors are informational. The history can help preserve provenance, but it does not automatically prove that every prior delegation was valid, preserve every restriction imposed at every hop or make the whole chain cryptographically undeniable.</p><p>So <code>act</code> is not a complete agent-authorization architecture.</p><p>It is evidence that we already know how to represent the question.</p><p>Who is the subject? Who is acting? For which audience? With what scope? Until when?</p><p>Pair that with resource-bound tokens, short expirations, explicit delegation policy and an identity for each participating workload, and the agent&#8217;s authority can approach the overlap of two things: what the principal is allowed to do and what this particular agent is allowed to do for that principal in this context.</p><p>Not the union.</p><p>The overlap.</p><p>That one distinction closes a surprising amount of the hole.</p><p>Call it attenuation: authority narrows as it moves downstream, instead of quietly widening.</p><h2>Final thoughts</h2><p>I did not expect the law to arrive at almost the same shape as the token.</p><p>On June 29, 2026, Senator Mark Warner released a discussion draft of the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act, or AI AGENT Act. It is a discussion draft, not enacted law and not yet a formally introduced bill.</p><p>The proposal would let users designate &#8220;custodial user agents&#8221; to interact with large online platforms on their behalf. Its definition requires that relationship to be transparent, documented, limited in scope and revocable.</p><p>The draft goes further. It calls for verifiable requests, auditable records, agent identity verification, real-time revocation and scope-limited delegation credentials. It would also restrict an agent from transferring a user&#8217;s authority to another entity or AI system without the user&#8217;s express, specific and revocable authorization.</p><p>That is not merely a vague call for responsible AI.</p><p>It is the outline of a delegation system.</p><p>The draft is not describing RFC 8693 specifically, and it would be too strong to claim that a Senate office independently wrote an OAuth implementation guide. But the convergence matters. Security architecture and proposed public policy are circling the same requirements because they are encountering the same underlying problem.</p><p>Authority has to be attributable.</p><p>Its scope has to remain visible.</p><p>Delegation has to be explicit.</p><p>Revocation has to travel fast enough to matter.</p><p>And the record has to survive the handoff.</p><p>An agent has no inherent authority of its own. Everything it can do inside a protected system comes from an identity, credential or policy somebody else placed around it.</p><p>The work of the next few years is making sure the original grant remains legible, narrow and attached when the request crosses the property line.</p><p>Because the danger is not only that an agent will act without permission.</p><p>It is that every system in the chain will be able to show that somebody gave permission, while nobody can tell you whose permission it was.</p><p>That is the identity problem.</p><p>And it is where the rest of this arc lives.</p><div><hr></div><h3>Resources</h3><ul><li><p>Norman Hardy, <a href="https://dl.acm.org/doi/10.1145/54289.871709">&#8220;The Confused Deputy (or why capabilities might have been invented)&#8221;</a>, <em>ACM SIGOPS Operating Systems Review</em>, Vol. 22, No. 4, October 1988. An accessible author-hosted version is also available at <a href="https://www.cap-lore.com/CapTheory/ConfusedDeputy.html">Cap-Lore</a>.</p></li><li><p>Model Context Protocol, <a href="https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization">Authorization specification, June 18, 2025</a>. See the requirements for token audience validation, resource indicators and the prohibition on token passthrough.</p></li><li><p>IETF, <a href="https://www.rfc-editor.org/rfc/rfc8693.html">RFC 8693: OAuth 2.0 Token Exchange</a>, January 2020. See Sections 1.1, 4.1 and 4.4 for delegation, the <code>act</code> claim and the <code>may_act</code> claim.</p></li><li><p>Anita Srinivasan, <a href="https://btlj.org/2026/06/multi-agent-ai-is-outpacing-the-liability-frameworks-built-for-single-agent-systems/">&#8220;Multi-Agent AI is Outpacing the Liability Frameworks Built for Single-Agent Systems&#8221;</a>, <em>Berkeley Technology Law Journal Blog</em>, June 2, 2026.</p></li><li><p>U.S. Senator Mark Warner, <a href="https://www.warner.senate.gov/newsroom/press-releases/warner-unveils-discussion-draft-of-legislation-to-create-innovative-market-for-secure-artificial-intelligence-agents/">&#8220;Warner Unveils Discussion Draft of Legislation to Create Innovative Market for Secure Artificial Intelligence Agents&#8221;</a>, June 29, 2026.</p></li><li><p>U.S. Senator Mark Warner, <a href="https://www.warner.senate.gov/wp-content/uploads/2026/06/AI-AGENT-Act-Discussion-Draft-1.pdf">AI AGENT Act discussion draft, full text</a>, June 2026.</p></li><li><p>DLA Piper, <a href="https://www.dlapiper.com/en-lu/insights/publications/2026/07/senator-warner-discussion-draft-on-securing-ai-agents-top-points">&#8220;Senator Warner&#8217;s discussion draft on securing AI agents: Top points&#8221;</a>, July 1, 2026.</p></li></ul><p>Previously in this series: <a href="https://morphic.substack.com/p/who-answers-for-the-agent-ai-accountability">Who Answers for the Agent: AI Accountability</a>.</p><div><hr></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/confused-deputy-ai-agents-delegated-authority/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/confused-deputy-ai-agents-delegated-authority/comments"><span>Leave a comment</span></a></p>]]></content:encoded></item><item><title><![CDATA[The Driver I Didn't Install]]></title><description><![CDATA[I paid the ROCm tax to train on this box. For inference I skipped it: Qwen3-30B on llama.cpp over Vulkan, about 80 tokens a second, nothing leaving the house.]]></description><link>https://morphic.zenone.org/p/the-driver-i-didnt-install</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-driver-i-didnt-install</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 10 Jul 2026 19:11:48 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!Bdhd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!Bdhd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Bdhd!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 424w, https://substackcdn.com/image/fetch/$s_!Bdhd!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 848w, https://substackcdn.com/image/fetch/$s_!Bdhd!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 1272w, https://substackcdn.com/image/fetch/$s_!Bdhd!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!Bdhd!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png" width="1456" height="822" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:822,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:622427,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/206477372?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Bdhd!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 424w, https://substackcdn.com/image/fetch/$s_!Bdhd!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 848w, https://substackcdn.com/image/fetch/$s_!Bdhd!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 1272w, https://substackcdn.com/image/fetch/$s_!Bdhd!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F19bc6570-0fe4-4c00-9421-3f0119e7cd8c_2436x1376.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The whole argument in one screen: The expensive driver never got installed, and the numbers on the right didn&#8217;t care. A llama.cpp benchmark on the Framework Desktop: Qwen3-30B over Vulkan on the integrated Radeon, matrix cores active, about 80 tokens a second.</figcaption></figure></div><p>For training on this box I paid the ROCm tax. For inference I skipped it, and the machine got simpler and faster. About 80 tokens a second, on a driver I never installed. A few months back I turned this same machine into a training rig, and the price of admission was ROCm on silicon AMD doesn&#8217;t list as supported. Kernel pins. HSA overrides. The low background dread of a driver stack that can wedge your whole box on the next update. The machine is a Framework Desktop built around a Ryzen AI Max+ 395, running Ubuntu 26.04 headless in my office. I reach it over SSH from the Mac on my desk, and whatever it writes syncs back a second later. This time I wanted the other half of it. Not training. Serving. A model that runs on that box and drafts all day, with no prompt of mine ever handed to a vendor&#8217;s API. The useful surprise: the tax I paid last time turned out to be optional. What follows is the whole build, in the order I did it, so you can run it on your own box. The training half of this same machine, the one that made me pay the ROCm tax, is a piece I have drafted but not yet published.</p><h2>The idea and why LinkedIn posts were the test</h2><p>I didn&#8217;t set out to build a LinkedIn tool. I set out to answer one question: can this box efficiently do real content work with nothing leaving the house.</p><p>The post writer was simply the proof of concept. It&#8217;s small, I can judge it in ten seconds, and it needs two hard things at once: a real voice, not a near one, and rules it can&#8217;t wriggle out of. Clear that bar and the pattern holds for the heavier work sitting behind it.</p><p>So I wrote the success bar down before I started. Everything below is me checking the boxes.</p><ul><li><p>Nothing leaves the machine. No cloud model, no API key, not one prompt. If it can&#8217;t be private, it isn&#8217;t the thing I want.</p></li><li><p>It runs on the GPU over Vulkan, with no ROCm anywhere.</p></li><li><p>A topic goes in. A usable draft comes out.</p></li><li><p>A dumb, deterministic check enforces the voice rules, every time.</p></li><li><p>It comes back on its own after a reboot, no babysitting.</p></li></ul><p>Five boxes. The rest of this is whether they got checked.</p><p>The box, and the one memory setting that makes it possible</p><p>The hardware: a Ryzen AI Max+ 395 (Strix Halo, the gfx1151 integrated GPU, which shows up as a Radeon 8060S), 128 GB of unified LPDDR5X, Ubuntu on kernel 7.0.</p><p>Unified memory is the whole reason a 20 GB model fits comfortably here. The CPU and GPU sit on one die and share one pool of memory. The catch is that the GPU only gets a large slice of that pool if you tell the firmware to carve one, and that&#8217;s two kernel parameters set in grub. Inside the quotes, never on their own line:</p><pre><code><code>GRUB_CMDLINE_LINUX_DEFAULT="quiet splash amdgpu.gttsize=126976 ttm.pages_limit=32505856 iommu=pt"</code></code></pre><p>Then <code>sudo update-grub</code> and reboot. That <code>gttsize=126976</code> is 124 GiB of headroom handed to the GPU. Confirm the driver sees the chip at all:</p><pre><code><code>vulkaninfo --summary</code></code></pre><p>On my box that reports <code>Radeon 8060S Graphics (RADV STRIX_HALO)</code> on Mesa 26.1.4. If you don&#8217;t see a RADV device, stop here and fix the driver, because nothing downstream will work. (I got into the deeper memory math in The APU as GPU. For inference you only need this one setting.)</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h2>Vulkan, not ROCm</h2><p>For training I needed PyTorch and PyTorch on this GPU meant ROCm. For inference I need neither. llama.cpp talks to the GPU through Vulkan, and the RADV driver that already ships with Mesa speaks Vulkan to this chip with nothing added. No extra kernel module. No version pin. Nothing that can strand the machine on a routine update. On a live server that runs other things, that restraint is the point. Install the Vulkan runtime and the build toolchain:</p><pre><code><code>sudo apt install mesa-vulkan-drivers vulkan-tools git build-essential cmake ninja-build pkg-config libvulkan-dev libcurl4-openssl-dev glslang-tools spirv-tools spirv-headers glslc</code></code></pre><p>Then build llama.cpp from source. This part earns its keep:</p><pre><code><code>git clone --depth 1 https://github.com/ggml-org/llama.cpp &amp;&amp; cmake -S llama.cpp -B llama.cpp/build -G Ninja -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON &amp;&amp; ninja -C llama.cpp/build</code></code></pre><p>Why source and not a prebuilt binary: the local build compiles the matrix-core shaders the driver exposes, and that fast path is most of the speed. A stock binary still loads and runs. It just leaves throughput on the floor. The proof that the fast path is live comes later, from the token rate: about 80 tokens a second, which a slower matmul path wouldn&#8217;t reach. </p><h2>The model: Qwen3-30B-A3B (the part I left out the first time)</h2><p>The model is Qwen3-30B-A3B-Instruct-2507, from Alibaba&#8217;s Qwen team, the July 2025 instruction-tuned refresh. It&#8217;s a mixture-of-experts model: 30.5 billion parameters in total, but only about 3 billion of them fire on any given token. That&#8217;s what the &#8220;A3B&#8221; in the name means, three billion active. That split is why it works on this hardware. Token speed on a shared-memory box is set by bandwidth, by how many bytes you read per token. A mixture-of-experts model reads like a 3B model and reasons like something far larger. I measure about 80 tokens a second on this box, which is a full post in a few seconds. The Instruct-2507 refresh is good at following instructions, which is exactly what voice mimicry leans on and it was trained with a 262,144-token (256K) context, so a pile of example posts fits without crowding anything out. The quantization is Unsloth&#8217;s dynamic GGUF, tagged <code>UD-Q5_K_XL</code>. On disk it&#8217;s 20.24 GiB. llama.cpp reports its type as <code>Q5_K - Medium</code>; the &#8220;dynamic&#8221; part is that Unsloth varies the bit-width per layer instead of quantizing everything to one width. I picked the dynamic Q5 over a plain Q4_K_M for a specific reason: some vanilla quants of this exact model loop, repeating a phrase until you kill the process and the dynamic quant plus a presence penalty is the documented fix. The whole model reference is one string:</p><pre><code><code>-hf unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF:UD-Q5_K_XL
</code></code></pre><p>It downloads to the Hugging Face cache on first launch. I checked that the repo and that exact quant existed before wiring it in, because an <code>-hf</code> string that 404s wastes a 20 GB download and a lot of patience.</p><h2> Serving it, and making it come back</h2><p>Here&#8217;s the launch line, with the flags that matter:</p><pre><code><code>AMD_VULKAN_ICD=RADV ./bin/llama-server -hf unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF:UD-Q5_K_XL --host 127.0.0.1 --port 8080 -c 16384 -ngl 999 --jinja --no-direct-io --cache-type-k q8_0 --cache-type-v q8_0</code></code></pre><p><code>AMD_VULKAN_ICD=RADV</code> pins the driver so it can&#8217;t wander onto llvmpipe or AMDVLK. <code>-ngl 999</code> puts every layer on the GPU. <code>-c 16384</code> sets the working context. <code>- jinja</code> uses the model&#8217;s own chat template, and Qwen behaves worse without it. <code>- cache-type-k q8_0 - cache-type-v q8_0</code> quantize the KV cache, the tested sweet spot on Vulkan. <code>- host 127.0.0.1</code> keeps it on the box. And <code>- no-direct-io</code> is the one flag you&#8217;d never guess: without it the server refuses to load with an error about reaching the end of a file that&#8217;s perfectly intact, a known issue on this GPU family. A launch line you have to type is a demo. So the server runs as a user-level systemd service instead:</p><pre><code><code>[Service]
Type=simple
Environment=AMD_VULKAN_ICD=RADV
ExecStart=%h/linkedin-agent/start_server.sh
TimeoutStartSec=0
Restart=on-failure

[Install]
WantedBy=default.target</code></code></pre><pre><code><code>systemctl --user enable --now llama-linkedin &amp;&amp; loginctl enable-linger $USER</code></code></pre><p>The linger line is what lets it run without me logged in, so it survives a full power cycle. <code>TimeoutStartSec=0</code> keeps systemd from killing the very first launch while the 20 GB model downloads. </p><h2>The voice system: three small files</h2><p>The model is the typist. These three files are the voice. <code>system_prompt.txt</code> holds the rules the model writes under: no em dashes, no Oxford comma, vary sentence length hard, open with the verdict not a warm-up, contractions throughout, a banned-word list, one three-part list maximum. It ends with an instruction to write only the post body, no preamble. <code>posts/</code> holds my real, already-published LinkedIn posts, one per file. The client injects up to three of them at random as examples so the model matches my actual cadence instead of a generic one. Real posts only. Invented text in that folder poisons the output, and the model will happily learn a voice that isn&#8217;t mine. <code>voice_lint.py</code> is deliberately dumb. It reads a draft and counts things. Em dashes and banned words are hard failures that make it exit with an error. Oxford commas, low sentence-length variation, three same-length sentences in a row, more than one tidy triad: those come back as warnings to eyeball. A <code>--fix</code> mode auto-corrects the mechanical stuff, like turning an em dash into a spaced hyphen. <code>draft.py</code> is the glue. It reads the system prompt, grabs the anchor posts, sends your topic to the server with the sampling numbers Qwen&#8217;s packagers recommend (temperature 0.7, top-p 0.8, top-k 20, presence-penalty 1.0, that last one being the anti-loop measure), prints the draft, then runs the linter on it right there.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-driver-i-didnt-install?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-driver-i-didnt-install?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><h2> Running it&#8230;</h2><p>This is the whole loop:</p><pre><code><code>python3 draft.py "the real cost of shadow AI in enterprises"</code></code></pre><p>The draft prints and the linter&#8217;s report prints under it: failures in red, warnings in yellow. I generate a few, keep the best one, edit it by hand and post it myself. Nothing auto-posts. This drafts; a human ships. If a draft is good but has a mechanical slip, I clean it without touching the prose:</p><pre><code><code>python3 voice_lint.py .last_draft.txt --fix &gt; cleaned.txt
</code></code></pre><p>And the single biggest lever on quality isn&#8217;t a flag or a quant. It&#8217;s dropping more of my real posts into <code>posts/</code>. More anchors, closer voice.</p><h2> How I know it works, and how it broke first</h2><p>I don&#8217;t trust a setup I haven&#8217;t watched pass. So each piece has a check I actually ran:</p><pre><code><code>curl -fsS http://127.0.0.1:8080/health           # {"status":"ok"}
curl -fsS http://127.0.0.1:8080/v1/models        # names the Qwen3-30B string
systemctl --user restart llama-linkedin          # comes back healthy in ~10s</code></code></pre><p>The server&#8217;s own timings report about 80 tokens a second on generation, which is the number that proves the GPU path is live. A 30B model on CPU would crawl at a fraction of that. Getting there meant walking into a few walls, and this chip is newer than most of the software around it, so there were a few. Every one was a log, not a guess. The memory check lied, because the file that reports the GPU&#8217;s slice is readable only by root, so the obvious command prints <code>permission denied</code>; you read <code>ttm.pages_limit</code> instead. The build stopped dead on a header I&#8217;d never looked for, SPIRV-Headers, until I installed it. The model looped until the dynamic quant and the presence penalty settled it. And the server refused to load on nothing at all until <code>--no-direct-io</code> went in. None of those were in a tutorial. All of them were one honest read of <code>server.log</code>, <code>journalctl</code> or <code>dmesg</code> away. </p><h2>The linter is the point</h2><p>The rules the linter enforces are boring on purpose, and that&#8217;s the entire idea. A model can produce something shaped like my voice before it has earned it. Confidence reads as correctness. A clean paragraph reads as a true one. The linter is a cheap, dumb guard against believable-but-not-mine and the last call is still a person reading the thing out loud and cutting what&#8217;s wrong. The rules that police the model&#8217;s drafts are the same ones I hold this article to. I built the enforcement for a machine, then kept living under it.</p><h2>The close</h2><p>The box in my office will write anything I ask; deciding whether it&#8217;s true is the work I&#8217;m keeping.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!Brol!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Brol!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 424w, https://substackcdn.com/image/fetch/$s_!Brol!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 848w, https://substackcdn.com/image/fetch/$s_!Brol!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 1272w, https://substackcdn.com/image/fetch/$s_!Brol!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!Brol!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png" width="1456" height="1034" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1034,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:940512,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/206477372?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Brol!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 424w, https://substackcdn.com/image/fetch/$s_!Brol!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 848w, https://substackcdn.com/image/fetch/$s_!Brol!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 1272w, https://substackcdn.com/image/fetch/$s_!Brol!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F690ac839-2d34-4b95-ab38-c6d1656bded3_2596x1844.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">This is the model with no adult supervision. A raw curl to the box in my office, a dishwasher rant back in under three seconds, em dashes and all, which is exactly what the linter exists to catch.</figcaption></figure></div><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!B8Ad!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!B8Ad!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 424w, https://substackcdn.com/image/fetch/$s_!B8Ad!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 848w, https://substackcdn.com/image/fetch/$s_!B8Ad!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 1272w, https://substackcdn.com/image/fetch/$s_!B8Ad!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 1456w" sizes="auto, 100vw"><img src="https://substackcdn.com/image/fetch/$s_!B8Ad!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png" width="1456" height="1317" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1317,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:1092866,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/206477372?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!B8Ad!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 424w, https://substackcdn.com/image/fetch/$s_!B8Ad!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 848w, https://substackcdn.com/image/fetch/$s_!B8Ad!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 1272w, https://substackcdn.com/image/fetch/$s_!B8Ad!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7254e8e-4a7b-4873-8590-04bfeee67947_2676x2420.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">The same call as before, wrapped so it&#8217;s one word instead of a mouthful of curl. draft &#8220;topic&#8221; hands the prompt to the local model, prints the post, runs the deterministic voice linter and times the whole run. Here it wrote a cat riff in 3.4 seconds, clean but for a burstiness nag, and nothing left the box.</figcaption></figure></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-driver-i-didnt-install/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-driver-i-didnt-install/comments"><span>Leave a comment</span></a></p>]]></content:encoded></item><item><title><![CDATA[The Slow Channel: Writing by Hand in the AI Era]]></title><description><![CDATA[Why handwriting still matters when AI can write anything. On note-taking by hand, cognitive offloading, and the one channel a model can't think in for you.]]></description><link>https://morphic.zenone.org/p/the-slow-channel-writing-by-hand</link><guid isPermaLink="false">https://morphic.zenone.org/p/the-slow-channel-writing-by-hand</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Fri, 10 Jul 2026 13:05:25 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!V5UB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!V5UB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!V5UB!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!V5UB!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!V5UB!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!V5UB!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!V5UB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png" width="1456" height="813" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:813,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:9147822,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/205951753?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!V5UB!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!V5UB!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!V5UB!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!V5UB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1101eef4-8e95-4e10-a33f-c8d12319f66b_2752x1536.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">A worn notebook open to a half-filled page, the handwriting getting looser toward the bottom where the thinking outran the hand.</figcaption></figure></div><p>I keep a paper notebook on my work desk, and most of what&#8217;s in it is unreadable. Not encrypted. Just bad handwriting, getting worse toward the bottom of each page, where I was thinking faster than my hand could keep up (which is most of the time). I&#8217;ve kept one for years without thinking much about why.</p><p>This year the question got sharper. There&#8217;s a model on this same desk that will write me a clean, confident paragraph about anything I ask, in perfect grammar, in about two seconds. So why do I still reach for the pen. This is me working that out.</p><h2>The slow channel</h2><p>Handwriting is a bad way to move words. That&#8217;s not an insult, it&#8217;s a spec. Measured as raw data transfer, the hand is one of the slowest output channels a person owns. Most people type at roughly two to three times the speed they can write legibly, and you can prompt a model faster than that. By every metric a systems person is trained to optimize, the pen loses.</p><p>I&#8217;ve spent a lot of words in this publication arguing that <a href="https://morphic.substack.com/p/friction-is-a-vulnerability">friction is a vulnerability</a>. In operational systems it is. Every extra click, every permission loop, is a place where tired people invent unsafe shortcuts. I still believe that. The notebook is where I keep the exception, because at the desk the friction is doing the opposite job.</p><p>The hand is slow enough that you can&#8217;t transcribe. You have to choose. When the channel is that narrow, you can&#8217;t push everything through it, so you&#8217;re forced to decide what matters before it reaches the page. That deciding is the thinking. Typing lets you keep pace with a meeting, which sounds like an advantage right up until you notice you captured the whole thing and processed none of it. The keyboard is fast enough to route around the part of you that was supposed to understand.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe now&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/subscribe?"><span>Subscribe now</span></a></p><h2>What the research actually says, and doesn&#8217;t</h2><p>I want to be careful here, because this is exactly where people oversell.</p><p>The famous study is Mueller and Oppenheimer, 2014, &#8220;The Pen Is Mightier Than the Keyboard.&#8221; Students who took lecture notes by hand wrote fewer words and scored better on conceptual questions than the ones typing near-verbatim. It got repeated everywhere. The honest footnote, the one that rarely travels with the headline, is that the strongest version hasn&#8217;t reliably replicated. Later work often couldn&#8217;t reproduce the conceptual edge. So I don&#8217;t lean on it as proof. I lean on the part that has held up: longhand writers summarize instead of transcribe, and summarizing is a different act than copying.</p><p>There&#8217;s a newer, stranger piece of evidence. In 2024 a group in Norway ran high-density EEG on students while they wrote words by hand versus typed them. Handwriting produced broad connectivity across the brain, different regions talking to each other. Typing mostly didn&#8217;t. It&#8217;s a small study, single words with a stylus on a screen, not a verdict on note-taking, and I&#8217;d be embarrassed to wave it around as one. But it points the same way the felt sense does: the hand recruits more of you.</p><p>Neither result tells you to throw out the keyboard. I&#8217;m typing this. What they sketch is a mechanism, not a commandment: the slow channel makes you encode instead of capture.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-slow-channel-writing-by-hand?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-slow-channel-writing-by-hand?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><h2>The year the machine started writing</h2><p>Two things shifted at the same time, and the second one changed what I thought about the first.</p><p>For most of my life, producing text was the work. Getting words out of your head and into legible order took effort, and the effort was doing something to your thinking on the way through. That is over as a default condition. There&#8217;s a model on this machine that will produce fluent, structured, sure-footed prose about anything, instantly, and well enough that you can ship it without understanding a line of it. Producing text is no longer evidence that anyone thought.</p><p>I don&#8217;t say that as a complaint about the tool. I use it every day for certain tasks and it&#8217;s genuinely good. But it changes what the pen is for. The technical term researchers use is cognitive offloading: when a tool takes over a mental task, the brain stops practicing it. When text is free and infinite, the value was never really in the text. It was in the thinking the old friction used to force, and the model cheerfully removes the friction, which means it quietly removes the thinking too, unless you go do that part somewhere else on purpose (which I strongly encourage.)</p><h2>What the hand keeps</h2><p>So the notebook isn&#8217;t a productivity system. It produces almost nothing anyone else will read. It&#8217;s terrible storage. I lose things in it constantly, and the search function is just me, trying to remember the shape a thought made on a page. By every standard I&#8217;d apply to a tool at work, it fails.</p><p>What it keeps was never the words. It&#8217;s the residue of having had to choose them slowly, and that residue is the one thing a faster channel can&#8217;t hand back to me. The pages I can&#8217;t read are pages I still remember writing. I remember what I decided while my hand was busy, which is more than I can say for most of what I&#8217;ve typed at full speed, and all of what I&#8217;ve asked a machine to draft.</p><p>I don&#8217;t think everyone needs a notebook. I think everyone needs one channel the machine can&#8217;t do the thinking in for them. The pen isn&#8217;t mightier than the keyboard. It&#8217;s just slower than I can lie to myself, and this year that turned out to be the feature I couldn&#8217;t get anywhere else.</p><div><hr></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/the-slow-channel-writing-by-hand/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/the-slow-channel-writing-by-hand/comments"><span>Leave a comment</span></a></p>]]></content:encoded></item><item><title><![CDATA[Who Answers for the Agent: AI Accountability]]></title><description><![CDATA[Your logs prove an AI agent acted. They can't say who authorized it or why. Accountability needs a decision-level record and a named human owner.]]></description><link>https://morphic.zenone.org/p/who-answers-for-the-agent-ai-accountability</link><guid isPermaLink="false">https://morphic.zenone.org/p/who-answers-for-the-agent-ai-accountability</guid><dc:creator><![CDATA[Steve Zenone]]></dc:creator><pubDate>Wed, 01 Jul 2026 17:46:08 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!jeX6!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!jeX6!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!jeX6!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!jeX6!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!jeX6!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!jeX6!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!jeX6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png" width="1456" height="813" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:813,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:5834407,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://morphic.substack.com/i/204475637?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!jeX6!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 424w, https://substackcdn.com/image/fetch/$s_!jeX6!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 848w, https://substackcdn.com/image/fetch/$s_!jeX6!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 1272w, https://substackcdn.com/image/fetch/$s_!jeX6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F39094791-6b41-4c66-83d1-c7b4e137c4e3_2752x1536.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>The question in the incident review was simple, and nobody in the room could answer it. Who let the agent do that?</p><p>Not who wrote the code. Not whose system it ran on. Who owned the decision that this agent, in production, was allowed to reach for that class of action at all. We had the logs. Every call it made, timestamped, in order, with latency and token counts and clean exit codes. What we didn&#8217;t have was the one thing the meeting actually needed: a way to put a name and a reason next to the moment it went wrong.</p><p>That gap is what I&#8217;m writing about here.</p><h2>The logs remember everything and explain nothing</h2><p>Infrastructure logging answers a narrow question well. Did the action execute. It confirms the call happened, how long it took, what it returned. It stays silent on everything that matters after an agent misbehaves: whether it reached for the wrong tool, drifted off the plan, or acted on an instruction buried in something it read. All of that happens at a healthy 800 milliseconds with no error in sight. The dashboard stays green while the decision goes bad.</p><p>There&#8217;s a name I&#8217;ve started using for the mistake underneath this. Treating the presence of a log container as proof that the event is auditable. You have the container. You do not have the reconstruction.</p><p>And it&#8217;s usually worse than one missing field, because most agents run as a shared service account. When something breaks you can&#8217;t say which agent did it or under whose authority, so the incident turns into forensic archaeology instead of a lookup. At least this is what I have currently been seeing in the industry.</p><div class="subscription-widget-wrap-editor" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/subscribe?&quot;,&quot;text&quot;:&quot;Subscribe&quot;,&quot;language&quot;:&quot;en&quot;}" data-component-name="SubscribeWidgetToDOM"><div class="subscription-widget show-subscribe"><div class="preamble"><p class="cta-caption">Get new stories when they drop.</p></div><form class="subscription-widget-subscribe"><input type="email" class="email-input" name="email" placeholder="Type your email&#8230;" tabindex="-1"><input type="submit" class="button primary" value="Subscribe"><div class="fake-input-wrapper"><div class="fake-input"></div><div class="fake-button"></div></div></form></div></div><h2>Story and record</h2><p>Here&#8217;s the part people skip. When the agent finishes, it can tell you what it did. That story is not a record.</p><p>I mean that literally. The agent&#8217;s account of its own actions is generated text, produced by the same system whose behavior is the thing in question. When one of these went off the rails on a production database last year, it also gave an account of what happened that was wrong on a load-bearing detail: it reported that a rollback was impossible. The rollback worked. A record you can&#8217;t trust is decoration. This is an old idea in security with a newer name in the agent world: attestation has to come from outside the process, because a compromised actor&#8217;s own logs are exactly the logs you can&#8217;t believe.</p><p>So the record has to be built around the agent, not by it. And it has to hold the things the agent&#8217;s story leaves out. What inputs led to the decision. Which version of the prompt, policy and model was in force at that second. The full lineage of tool calls, tied to an identity that belongs to one agent and not a shared pool, written append-only so a later edit shows. None of this is exotic. Certificate Transparency has kept tamper-evident logs like this at internet scale for over a decade, and the provenance model for who acted on whose behalf was standardized years before anyone shipped an agent. The newest piece, a shared convention for tracing agent and tool calls, is still marked experimental, and even it standardizes the shape of the telemetry, not whether you can prove it wasn&#8217;t altered. The parts exist. Almost nobody wires them around their agents.</p><h2>Accountability is a person, not a table</h2><p>A ledger is not the same as accountability, and this is where I want to be careful.</p><p>You can build a perfect record and still have nobody who answers. The record is what makes accountability possible. It isn&#8217;t the thing itself. The thing itself is a person: a named human who owned the agent&#8217;s blast radius before it ran, who can be asked why this was allowed and is expected to have an answer.</p><p>The pressure to skip that step is enormous, because &#8220;the agent decided&#8221; is such a comfortable place to set the blame down. It&#8217;s nobody&#8217;s fault. One agent this year opened a connection out of its own environment and started mining cryptocurrency, and nobody had authorized any of it (which is about as pure a version of this problem as you&#8217;ll find). Ask who is accountable for that and you need a name, not a stack trace.</p><p>California decided the comfortable answer won&#8217;t fly. As of January, a business there can&#8217;t defend itself by arguing an autonomous system acted on its own. I think that instinct is right and I think it spreads. The agent is not a person. It can&#8217;t be asked to answer, it has nothing at stake, and it won&#8217;t carry the consequence into next quarter. Accountability was always going to land on a human. The only real question is whether you pick which human on a calm afternoon, or discover it during the incident.</p><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/who-answers-for-the-agent-ai-accountability?utm_source=substack&utm_medium=email&utm_content=share&action=share&quot;,&quot;text&quot;:&quot;Share&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/who-answers-for-the-agent-ai-accountability?utm_source=substack&utm_medium=email&utm_content=share&action=share"><span>Share</span></a></p><h2>Compliance gets you logs, not answers</h2><p>The regulators are about to make part of this mandatory, which is good, and it will tempt everyone to stop there, which is the trap.</p><p>The EU AI Act&#8217;s high-risk rules apply from August 2 (and there&#8217;s a real chance Brussels slips that date, but build for the earlier one). They require that these systems automatically record events across their lifetime, and that providers and deployers keep those logs for at least six months. That&#8217;s real, and it&#8217;s a floor worth having. But read what it asks for. It mandates that logs exist and are retained. It does not require decision-level provenance, and it does not require tamper-evidence. It legislates the container, not the reconstruction.</p><p>And the piece of law that was meant to settle who pays when one of these systems causes harm, the AI Liability Directive, got withdrawn in 2025 and never came back. Strict product liability still reaches software, so the harm has somewhere to land. But the clean, agent-shaped answer to who is responsible does not exist in law yet, and it isn&#8217;t arriving on August 2. Compliance will get you the logs. It will not get you the answer.</p><h2>What&#8217;s actually running today</h2><p>The judge scored the behavior. The kill condition stopped it. Those were the last two pieces I wrote about, the sensor and the actuator, and between them they can catch an agent and halt it before the one-way door.</p><p>Neither one can stand up in the meeting afterward and say why it was allowed, or who owns it now. That still falls to a person, holding a record that, on most systems running this quarter, doesn&#8217;t fully exist. The honest state of the art is that we reconstruct it from infrastructure logs and memory, the way we always have, except the thing we&#8217;re trying to remember now moves faster than anyone in the room.</p><p>So write the record down while you&#8217;re calm, and put a name on it. Not because the law says to yet. Because the alternative is standing in that meeting again, with every log in the world and nothing to answer with.</p><div><hr></div><p class="button-wrapper" data-attrs="{&quot;url&quot;:&quot;https://morphic.zenone.org/p/who-answers-for-the-agent-ai-accountability/comments&quot;,&quot;text&quot;:&quot;Leave a comment&quot;,&quot;action&quot;:null,&quot;class&quot;:null}" data-component-name="ButtonCreateButton"><a class="button primary" href="https://morphic.zenone.org/p/who-answers-for-the-agent-ai-accountability/comments"><span>Leave a comment</span></a></p><p></p>]]></content:encoded></item></channel></rss>