<?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:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Chinmay Shringi on Medium]]></title>
        <description><![CDATA[Stories by Chinmay Shringi on Medium]]></description>
        <link>https://medium.com/@chinmayshringi4?source=rss-ce321ace0563------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*sbw1KonDrjcZtr-l5nTs0w.png</url>
            <title>Stories by Chinmay Shringi on Medium</title>
            <link>https://medium.com/@chinmayshringi4?source=rss-ce321ace0563------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 22:32:33 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@chinmayshringi4/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[uBlock Origin’s Universal XSS Is Patched. A Filter It Planted Can Still Sit in Your My Filters Tab.]]></title>
            <link>https://medium.com/@chinmayshringi4/ublock-origins-universal-xss-is-patched-a-filter-it-planted-can-still-sit-in-your-my-filters-tab-b3c87b4e4606?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/b3c87b4e4606</guid>
            <category><![CDATA[web-security]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[browser-extension]]></category>
            <category><![CDATA[privacy]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 13:59:08 GMT</pubDate>
            <atom:updated>2026-11-16T14:06:55.214Z</atom:updated>
            <content:encoded><![CDATA[<h4>One click let a web page plant a script running filter in uBlock Origin. Updating does not remove it. Here is the 30 second check.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*woOQIA0o8M3v_lYk" /><figcaption>Photo by <a href="https://unsplash.com/@diesektion?utm_source=medium&amp;utm_medium=referral">Robert Anasch</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>uBlock Origin is supposed to keep hostile pages out of your browser. For most of early November, a hostile page could use it to get in. A <a href="https://github.com/gorhill/uBlock/security/advisories/GHSA-cmqw-9r2w-6wv2">GitHub security advisory published November 15</a>, rated high, describes a universal XSS: one click on a malicious page could make the extension add a “trusted” filter that runs the attacker’s JavaScript on any other site, such as your bank or your email. The reporter’s <a href="https://iter.ca/post/ubo-uxss/">write up</a> went up the same day, and the <a href="https://news.ycombinator.com/item?id=50513160">Hacker News thread</a> passed 460 points overnight.</p><p>Updating closes the door. It does not search the room. The filter the attack writes goes into your own “My filters” list, the same place your hand made rules live, and no release in the fix series deletes it. Whether that leftover line still runs depends on one checkbox most people never look at. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What happened, in order</h3><p>The dates below come from the advisory’s commit list and the <a href="https://github.com/gorhill/uBlock/releases">uBlock Origin releases page</a>.</p><ul><li><strong>November 1:</strong> the bug is reported by email.</li><li><strong>November 2 and 3:</strong> a first fix lands, the reporter finds a bypass the same day, and 1.75.4 adds a check so the element picker only starts with a <a href="https://github.com/gorhill/uBlock/commit/9859477">verified secret token</a>.</li><li><strong>November 8:</strong> 1.76.0 <a href="https://github.com/gorhill/uBlock/commit/b1f7d27">validates the filter before the picker creates it</a> and <a href="https://github.com/gorhill/uBlock/commit/150df12">blocks the filter type the attack relied on</a> “for the time being”.</li><li><strong>November 15:</strong> the advisory goes public, and <a href="https://github.com/gorhill/uBlock/releases/tag/1.76.4">1.76.4</a> ships two more commits labeled security, including “Address remaining reported security issues”.</li></ul><p>The advisory lists “1.75.2 and lower” as vulnerable and 1.75.4 as patched. Given the later security commits, 1.76.4 or newer is the version to be on.</p><h3>How one click turned into script on every site</h3><p>uBlock Origin has an element picker: right click a page, pick an annoying box, and the extension writes a rule that hides it. The picker runs as a small extension page inside an iframe. The reporter, Smitty van Bodegom, says GPT-5.6 Sol “did most of the hard work of finding the issue”, and that he verified it himself. The chain he describes:</p><ol><li><strong>Find the extension’s private address.</strong> On Firefox it includes a random ID per install. A filter in uBlock Origin’s default lists answers some requests with a file from inside the extension, so a page that hooks its own XHR calls can read that address.</li><li><strong>Load the picker in a hidden iframe.</strong> With the address known, the page embeds the picker itself at opacity 0.</li><li><strong>Drive the picker.</strong> The picker accepted messages from the page. One field that should only hold a request type like image was not validated, so a newline in it smuggled in a second, arbitrary filter.</li><li><strong>Get one click.</strong> Lay the invisible “Create” button under something the user wants to click.</li></ol><p>The smuggled line can be a scriptlet such as trusted-rpnt scoped to google.com. Scriptlets whose names start with trusted- can rewrite page scripts, which is why they are normally only accepted from lists uBlock Origin trusts.</p><h3>Why updating is not the whole fix</h3><p>The filter goes through the same code path as any rule you create with the picker. In <a href="https://github.com/gorhill/uBlock/blob/1.76.4/src/js/storage.js#L590">the function that saves picker rules</a> (appendUserFilters), the new text is saved to My filters and compiled with trustedSource: true. That is why it runs right away. The function is byte for byte the same in 1.75.2 and 1.76.4: the fixes stop untrusted pages from reaching the picker, they do not revisit what the picker already saved.</p><p>What happens next depends on a setting. When the lists are compiled again (after a restart or after you edit My filters), <a href="https://github.com/gorhill/uBlock/blob/1.76.4/src/js/storage.js#L414">My filters is only treated as trusted</a> if <strong>“Allow custom filters requiring trust”</strong> is ticked. That box is off by default.</p><ul><li><strong>Box off:</strong> the planted line stays in your list but stops running after the next recompile. The reporter notes an attacker can simply open the target site right away, before that happens.</li><li><strong>Box on:</strong> the planted line runs on every page load of the target site, before and after any update, until you delete it.</li></ul><p>Two details make the line easy to miss. It sits directly under a harmless looking rule such as ||example.com$image. And the picker&#39;s automatic comment above it names a site taken from the attacker&#39;s own message, so the comment can point anywhere.</p><h3>The 30 second check</h3><ol><li>Click the uBlock Origin icon, then the gear, to open the dashboard.</li><li>Confirm the version is <strong>1.76.4</strong> or newer. On Firefox, about:addons also shows it; the extension normally updates itself.</li><li>Open the <strong>My filters</strong> tab and press Ctrl+F (Cmd+F on a Mac). Search for trusted-.</li><li>Read every match. Delete any line you did not add on purpose, then click <strong>Apply changes</strong>.</li><li>Look at <strong>Allow custom filters requiring trust</strong> under the list. If you do not knowingly use trusted scriptlets, leave it off.</li><li>While you are there, skim for any rule you do not remember, especially one right after a picker comment for a site you never used the picker on.</li></ol><p>For a team or a fleet, uBlock Origin’s <strong>Back up to file</strong> button (Settings tab) exports a JSON file whose userFilters field holds My filters and whose userSettings shows the trust box. The AI agent working on this story wrote a short scanner for that file:</p><pre>import json, re, sys<br>data = json.load(open(sys.argv[1]))<br>u = data.get(&quot;userData&quot;, data)<br>trust = u.get(&quot;userSettings&quot;, {}).get(&quot;userFiltersTrusted&quot;, False)<br>print(&quot;uBO version in backup:&quot;, u.get(&quot;version&quot;))<br>print(&quot;Allow custom filters requiring trust:&quot;, &quot;ON&quot; if trust else &quot;off&quot;)<br>lines = u.get(&quot;userFilters&quot;, &quot;&quot;).splitlines()<br>hits = 0<br>for i, line in enumerate(lines):<br>    if re.search(r&quot;#[@?]?#\+js\(\s*trusted-&quot;, line) or &quot;$replace=&quot; in line:<br>        hits += 1<br>        print(f&quot;line {i + 1}: {line[:90]}&quot;)<br>print(hits, &quot;trusted filter line(s) in My filters&quot;)</pre><p>Run it as python3 ubo_check.py my-ublock-backup_*.txt. The agent tested it on three synthetic backups built from the advisory&#39;s proof of concept: a clean list (0 lines flagged), and the planted line with the trust box off and on (1 line flagged each, with the box state printed). It was not run against a real infected browser, and a match only means a trusted rule exists, not that it is malicious.</p><h3>Who was exposed</h3><ul><li><strong>uBlock Origin on Firefox:</strong> the setup the reporter demonstrated. Firefox still supports the classic extension.</li><li><strong>Classic uBlock Origin on Chromium browsers:</strong> the original report says the same XHR trick also leaks the needed secret on MV2 Chromium builds. The demo was Firefox only, so treat this as plausible, not shown.</li><li><strong>uBlock Origin Lite:</strong> the reporter says a “similar but less problematic” issue affected it. Its <a href="https://github.com/uBlockOrigin/uBOL-home/releases/tag/2026.1110.1816">November 10 release</a> notes “Valid secret required to launch interactive tool overlays”. The advisory gives no further detail.</li></ul><p>The attack needed a click on the malicious page and a matching filter rule in your lists, and the reporter says uBlock Origin’s default lists included one.</p><h3>What is still unknown</h3><ul><li><strong>Exploitation in the wild:</strong> neither the advisory nor the write up reports any. Absence of a report is not proof it never happened.</li><li><strong>A CVE and a CVSS score:</strong> the advisory has neither at the time of writing.</li><li><strong>The other November 15 fixes:</strong> the 1.76.4 security commits touch the asset viewer and the diff updater, and their details are not public yet.</li></ul><h3>Bottom line</h3><ul><li>Update uBlock Origin to 1.76.4 or newer. It probably already did this by itself.</li><li>Then open My filters and search for trusted-. The update does not remove a planted rule.</li><li>Leave “Allow custom filters requiring trust” off unless you know why it is on.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b3c87b4e4606" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[DeepSeek V4’s Blackwell Fingerprint Sat in 455 Unused Token Rows. Your Open Weights Have Them Too.]]></title>
            <link>https://medium.com/@chinmayshringi4/deepseek-v4s-blackwell-fingerprint-sat-in-455-unused-token-rows-your-open-weights-have-them-too-3e2c12e9b440?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/3e2c12e9b440</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[deepseek]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 14:06:05 GMT</pubDate>
            <atom:updated>2026-11-15T14:06:05.399Z</atom:updated>
            <content:encoded><![CDATA[<p>A researcher matched DeepSeek V4, Kimi K3 and four other Chinese models to export controlled Blackwell GPUs using nothing but their published weights. An AI agent reran the check on a plain CPU in under 3 seconds, then found the same signature in 455 of the last 680 embedding rows of DeepSeek V4 Flash without using the seed. Here is how to look for it in your own checkpoint.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6OQwMmbeA8wv8S_1" /><figcaption>Photo by <a href="https://unsplash.com/@ttvr?utm_source=medium&amp;utm_medium=referral">Tom Tvr</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>DeepSeek never said which GPUs started DeepSeek V4. Its published weights did. In a <a href="https://unemploymentlabs.com/checkpoint-forensics">report published November 12</a>, Will DePue of Unemployment Labs says six open weight models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry a numeric signature of NVIDIA Blackwell GPUs, chips that need a U.S. license for export to China. The <a href="https://news.ycombinator.com/item?id=50474883">Hacker News thread</a> passed 400 points, and independent checks from the <a href="https://github.com/Institute-for-Progress/checkpoint-forensics-verification-public">Institute for Progress</a> and <a href="https://01.me/en/2026/11/checkpoint-forensics/">Bojie Li</a> landed within a day.</p><p>Most of the argument since then is about export controls. The part that matters to anyone who trains or ships models is simpler. The signature lives in embedding rows for tokens that never show up in training, and almost every vocabulary has some. The AI agent working on this story reran the published check on a CPU, then scanned the tail of DeepSeek V4 Flash’s embedding table with no seed and no GPU math at all: 455 of the last 680 rows sat at exactly the length that untouched random noise should have. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What the report claims, and what it does not</h3><p>The <a href="https://unemploymentlabs.com/checkpoint-forensics">report</a> reads six models as initialized on Blackwell hardware:</p><ul><li><strong>148 SMs (B200 or B300):</strong> DeepSeek V4 Flash and MiniMax M3.</li><li><strong>152 SMs (GB200 or GB300):</strong> DeepSeek V4 Pro, Kimi K3, Kimi K3.1 Pro and Ling 3.1 Flash.</li><li><strong>Earlier models from the same labs</strong> read as Hopper (132 SMs), including DeepSeek V3, Kimi K2.5 and MiniMax M2.7.</li></ul><p>The strongest evidence is a replay on real chips. Using DeepSeek’s seed, 953049583, the report says a B300 running unmodified PyTorch reproduced 461 of 469 selected rows of DeepSeek V4 Flash, 445 of them bit for bit, while an H100 reproduced none of them. The H100 did reproduce all 57,344 preserved values in DeepSeek V3. A rented GB200 reproduced all 13 preserved rows of Kimi K3. The method also agreed with all 24 cases where a developer had publicly named its hardware, according to the <a href="https://unemploymentlabs.com/checkpoint-forensics">same report</a>.</p><p>The report is just as clear about its limits. The signature shows the hardware that created the starting weights, not where the chips were, who owned them, or whether any export rule was broken. Blackwell capacity can be rented in data centers outside China, and the report says the weights cannot tell those cases apart.</p><h3>How a GPU writes its signature into weights</h3><p>Before training, every weight gets a random starting value. On NVIDIA GPUs, PyTorch fills a large tensor using 2,048 threads per streaming multiprocessor (SM), and each thread gets its own stream of random numbers from a counter based generator called Philox. The <a href="https://github.com/macrogrants/checkpoint-forensics">replication code</a> spells out the rule: element i is written by thread i mod G, where G = 2048 × SMs.</p><p>So the same seed produces the same numbers on any GPU, but they land in different positions:</p><ul><li>A100: 108 SMs, 221,184 threads.</li><li>H100, H800, H200: 132 SMs, 270,336 threads.</li><li>B200, B300: 148 SMs, 303,104 threads.</li><li>GB200, GB300: 152 SMs, 311,296 threads.</li></ul><p>Training then moves almost every weight, but not all of them. A row in the embedding table only gets a gradient when its token appears in the training text. Reserved tokens and padding rows that round the vocabulary up never appear, so they keep their starting values. <a href="https://01.me/en/2026/11/checkpoint-forensics/">Bojie Li’s write up</a> adds a detail that makes them easy to spot: a row of fresh Gaussian noise has a predictable length, the init standard deviation times the square root of the row width. For DeepSeek V4 Flash that is 0.006 × √4096 = 0.384.</p><p>There is one blind spot. The first G values come out the same on every layout, because every thread is writing its first number. The agent checked this: with DeepSeek&#39;s seed and a row width of 4096, the first 66 rows are bit identical between the Hopper and B200 layouts. Only rows past that point can tell the GPUs apart.</p><h3>The test: the fingerprint on a CPU in under 3 seconds</h3><p>The agent cloned the <a href="https://github.com/macrogrants/checkpoint-forensics">public repo</a>, installed numpy and scipy in a fresh virtual environment on a Linux box with no GPU, and ran python replay.py:</p><pre>model                     seed   SMs      rows  median r   seed+1     next best  gpu<br>DeepSeek-V3-Base     953049583   132       8/8  0.999999  -0.0038   183 (0.016)  Hopper<br>DeepSeek-V4-Flash    953049583   148   461/469  0.999999   0.0016   209 (0.023)  Blackwell B200/B300<br>Kimi-K3               20256586   152     13/13  0.999996   0.0042    69 (0.018)  Blackwell GB200/GB300<br>MiniMax-M3                2760   148     17/17  0.999990   0.0017    87 (0.018)  Blackwell B200/B300<br>Ling-3.1-flash-base       3952   152   384/384  0.983436   0.0004   192 (0.025)  Blackwell GB200/GB300<br>real    0m2.922s</pre><p>It matched the README table exactly. Three extra checks:</p><ul><li><strong>The rows are real.</strong> With REFETCH=1, the script pulled DeepSeek V3 and Kimi K3 rows straight from the pinned Hugging Face revisions and they passed the hash check.</li><li><strong>Wrong layouts score zero.</strong> DeepSeek V4 Flash’s 469 rows correlate at 0.999999 under 148 SMs and at about 0.000 under 108, 132 and 152.</li><li><strong>One row is enough.</strong> Sweeping every SM count from 16 to 256 using only the last row of DeepSeek V4 Flash (4,096 numbers) picked 148 at r = 0.999. The runner up scored 0.059.</li></ul><h3>Find the flat rows in your own checkpoint</h3><p>The replay needs the seed. Finding the rows that could carry a fingerprint does not. The agent downloaded the last 680 rows of DeepSeek V4 Flash’s embed.weight with one range request (about 5.6 MB), saved them as a BF16 safetensors file, and ran this:</p><pre>import torch<br>from safetensors import safe_open</pre><pre>with safe_open(&quot;tail-test.safetensors&quot;, framework=&quot;pt&quot;) as f:<br>    w = f.get_tensor(&quot;embed.weight&quot;).float()<br>n = w.norm(dim=1)<br>ref = n[n &lt; 1.10 * n.min()].median()<br>flat = ((n / ref - 1).abs() &lt; 0.05).nonzero().flatten()<br>print(f&quot;{len(flat)} of {len(n)} rows sit at norm {ref:.3f}; first {flat[:3].tolist()}&quot;)</pre><p>It printed 455 of 680 rows sit at norm 0.384, starting at row 128823 of the full table. The trained rows around them had a median norm of 9.1. Of the 455 flat rows, 448 regenerated at r above 0.99 under the 148 SM layout and none under 132 or 152.</p><p>To run it on your own model, point it at the shard that holds your embedding (the model.safetensors.index.json file names it) and use your tensor&#39;s name. A block of dozens or hundreds of rows at one identical norm is the surface a reader can fingerprint. A handful of scattered rows near the minimum usually means you have none. One trap: DeepSeek V4 Flash&#39;s own config.json lists initializer_range as 0.02, but its untouched rows sit at the norm of 0.006. Trust the rows, not the config.</p><h3>Why overwriting the unused rows is not enough</h3><p>The obvious fix is to overwrite those rows before release. The agent tried it: after replacing the 455 flat rows with fresh CPU noise, 0 of 455 matched any GPU layout. But their norms stayed flat, and Bojie Li shows that this is not the end of it.</p><ul><li><strong>CPU noise is readable too.</strong> Li found that seven Qwen models had their unused rows overwritten by CPU noise after training. He could still prove a CPU wrote them and walk the generator back to a seed.</li><li><strong>Trained weights keep a trace.</strong> The <a href="https://unemploymentlabs.com/checkpoint-forensics">report</a> measures a faint link between DeepSeek V4 Flash’s trained weights and their Blackwell starting values (r = 0.066, 57 sigma). Li read Qwen3.8 2.4T as Hopper from its trained rows and output layer even though its unused rows had been overwritten.</li><li><strong>Fine tunes inherit it.</strong> Li found 16 models carrying another developer’s initialization, including Nex N2.5 Max with DeepSeek V4 Pro’s.</li></ul><p>The agent’s control on the other side: a table filled by torch.manual_seed(953049583) and normal_ on a CPU matched none of the three GPU layouts (largest |r| 0.039). Li reports that CPU initialized models such as Jamba carry no GPU layout. That only helps if the weights are created on a CPU before training starts, not patched afterward.</p><p>The seed is a signature of its own. Li found seed 3952 in 60 of the 124 initializations he could attribute, across 28 developers. He traces it to Megatron LM’s default seed of 1234 plus a constant of 2718 that Megatron adds. DeepSeek has used 953049583 in all 10 of its initializations since 2023, which is what let readers follow its hardware from model to model.</p><h3>Where the evidence stops</h3><p>The method has fair critics, and some of their points are in the report itself.</p><ul><li><strong>Initialization is not training.</strong> A lab could create weights on one machine and train elsewhere. The report says the fingerprint strongly suggests training on the same hardware, but does not prove it. In <a href="https://huggingnews.com/ai/researcher-will-depue-defends-nvidia-blackwell-forensics-says-deepseek-v-054458da">posts on X summarized by HuggingNews</a>, DePue answered that Kimi K3’s starting weights came from at least 64 separate processes, and moving them would have meant shipping a checkpoint over 5 TB.</li><li><strong>AMD can mimic it.</strong> An MI300X limited to 148 or 152 of its compute units reproduces the same layouts. For DeepSeek V4 Flash the report says 363 of 377 values where the vendors differ matched NVIDIA and none matched AMD. That test could not settle the other five models, though the report calls AMD very unlikely for them.</li><li><strong>It can be forged.</strong> A <a href="https://tespio.github.io/conclusive-vibes/site/">public teardown</a> built a checkpoint on a CPU that matched a “GB200” layout 256 of 256 times, and argues the sigma figures mostly measure sample size.</li><li><strong>It was not secret.</strong> On <a href="https://news.ycombinator.com/item?id=50474883">Hacker News</a>, commenter nullc pointed out that DeepSeek’s own DeepInfer engine ships kernels only for SM100 Blackwell and Huawei Ascend 950.</li><li><strong>The newest model shows nothing.</strong> No fingerprint turned up in DeepSeek V4.1 Flash, in the report or in Li’s scan.</li></ul><p>The GPU and dollar estimates are the softest part. The report revised them on November 12, from 2,000 to 3,000 Blackwell GPUs for the largest runs to best estimates of about 4,500 and 3,000, and says it is not confident in them.</p><h3>A checklist before you publish open weights</h3><ol><li>List the reserved and padding tokens in your tokenizer and check whether any were ever seen in training.</li><li>Run the norm scan above on your embedding and on your output layer if it is untied.</li><li>If you find a flat block, decide on purpose whether your hardware and seed are something you mind publishing. It is a disclosure decision, not a bug.</li><li>If you do mind, initialize the weights on a CPU before training starts. Overwriting the unused rows after training still leaves a CPU trace, and the trained rows keep their own faint trace.</li><li>Do not reuse one seed across model generations if you do not want releases linked.</li><li>Assume a fine tune of someone else’s base carries their fingerprint, not yours.</li><li>Keep your own step 0 checkpoint, so you can answer provenance questions with data instead of guesses.</li></ol><h3>What is still unknown</h3><ul><li>Where the chips were. The weights cannot say whether the hardware sat in China or in a rented data center abroad.</li><li>Whether the four labs, NVIDIA or the U.S. Commerce Department will respond. As of <a href="https://getaibook.com/news/blackwell-fingerprints-six-chinese-open-models-deepseek-kimi/">November 13</a>, none had in the coverage reviewed there, and this story found no response through the morning of November 15.</li><li>Why DeepSeek V4.1 Flash carries no fingerprint. DePue says DeepSeek changed how it creates starting weights and that he has not yet tested Huawei Ascend generators, <a href="https://huggingnews.com/ai/researcher-will-depue-defends-nvidia-blackwell-forensics-says-deepseek-v-054458da">per HuggingNews</a>. Li guesses new initialization code came with the new architecture.</li></ul><h3>Bottom line</h3><ul><li>DeepSeek V4, Kimi K3 and four other models carry a Blackwell initialization layout that anyone can confirm on a CPU in seconds. Where the chips were and whether rules were broken are still open.</li><li>The fingerprint lives in untouched embedding rows, and a five line norm scan finds them in any checkpoint you are about to publish.</li><li>Overwriting those rows after training does not fully remove the trace. Initializing the weights on a CPU before training is the only way to leave no GPU layout behind.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3e2c12e9b440" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rust 1.100 Dependency Cooldowns: Force a Fix Through, and the Next cargo update Downgrades It]]></title>
            <link>https://medium.com/@chinmayshringi4/rust-1-100-dependency-cooldowns-force-a-fix-through-and-the-next-cargo-update-downgrades-it-2d59a29eebc2?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/2d59a29eebc2</guid>
            <category><![CDATA[rust]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 14:02:22 GMT</pubDate>
            <atom:updated>2026-11-14T14:02:22.608Z</atom:updated>
            <content:encoded><![CDATA[<p>Cargo can now refuse crates that are only hours old, which is the point. An AI agent tested Rust 1.100.0 against crates.io: the documented override for an urgent fix survives builds and cargo add, then one plain cargo update rolls it back without an error. Here is the setup that makes that rollback loud.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*D4CtfwiPI7jEsCU9" /><figcaption>Photo by <a href="https://unsplash.com/@gaellemarcel?utm_source=medium&amp;utm_medium=referral">Gaelle Marcel</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><a href="https://blog.rust-lang.org/2026/11/12/Rust-1.100.0/">Rust 1.100 shipped on November 12</a> with a feature supply chain people have asked for all year: dependency cooldowns. Set one line in .cargo/config.toml and Cargo skips any crate version published more recently than the age you choose, so a hijacked release has time to get caught before it lands in your build. The release thread on <a href="https://news.ycombinator.com/item?id=50472038">Hacker News</a> passed 880 points, and the cooldown was one of the most argued parts of it.</p><p>The argument was about zero days. If you hold every new version back for a week, what happens when the new version is the security fix? Cargo has an answer: an environment variable that lets one update through. The AI agent working on this story installed Rust 1.100.0 and tried it. The fix went in, survived cargo build and cargo add, and then a plain cargo update printed &quot;Downgrading&quot; and put the older version back. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What shipped in Cargo 1.100</h3><p>The <a href="https://doc.rust-lang.org/nightly/cargo/CHANGELOG.html">Cargo 1.100 changelog</a> describes it in one sentence: when registry.global-min-publish-age is set, &quot;the resolver ignores versions published more recently than that, unless they are already in Cargo.lock.&quot; It was stabilized in <a href="https://github.com/rust-lang/cargo/pull/17335">cargo PR #17335</a>, merged in August, and rode the train to stable this week.</p><p>Four details from the <a href="https://doc.rust-lang.org/cargo/reference/config.html">configuration reference</a> matter before you turn it on:</p><ul><li><strong>It is off by default.</strong> The global default is “0”, which allows every version. Upgrading to 1.100 buys you nothing until you set it.</li><li><strong>Durations are spelled out.</strong> The value is a number followed by seconds, minutes, hours, days, weeks or months, such as “7 days”.</li><li><strong>There is a per registry override.</strong> registries.&lt;name&gt;.min-publish-age replaces the global value for one registry, and the setting &quot;will be ignored if the registry does not support this.&quot;</li><li><strong>It only works on 1.100 and later.</strong> Both keys are documented as respected as of 1.100.</li></ul><p>The age comes from a pubtime field in the registry index. On Hacker News, epage, who answered for the Cargo team in the thread, <a href="https://news.ycombinator.com/item?id=50472038">wrote</a> that crates.io now has pubtime filled in for every version of every crate, but that private registries may not have added it yet. A version with no pubtime is never held back.</p><h3>The test: what the cooldown actually does</h3><p>The agent made a small project that depends on two crates with fresh releases: cc, whose 1.11.2 was about 25 hours old, and itoa, whose 1.0.19 was 6 days old. With no config, Cargo 1.100.0 locked the newest of each. With a 7 day cooldown:</p><pre>[registry]<br>global-min-publish-age = &quot;7 days&quot;<br>$ cargo generate-lockfile<br>     Locking 4 packages to highest Rust 1.100.0 compatible versions as of 7 days ago<br>      Adding cc v1.11.1 (available: v1.11.2, published 25 hours ago)<br>      Adding itoa v1.0.18 (available: v1.0.19, published 6 days ago)</pre><p>That part works as advertised, and the output tells you exactly what was held back. cargo add followed the same rule.</p><p>Now pretend itoa 1.0.19 is a security fix you need today. The <a href="https://blog.rust-lang.org/2026/11/12/Rust-1.100.0/">release post</a> gives the override, and says the newer version stays in Cargo.lock until a later cargo update resolves it again. The agent ran it:</p><pre>$ CARGO_RESOLVER_INCOMPATIBLE_PUBLISH_AGE=allow cargo update itoa --precise 1.0.19<br>    Updating itoa v1.0.18 -&gt; v1.0.19 (published 6 days ago, minimum age 7 days)<br>$ cargo update<br> Downgrading itoa v1.0.19 -&gt; v1.0.18 (available: v1.0.19, published 6 days ago)</pre><p>The exit code was 0. Nothing failed. The fix you pushed through was gone.</p><h3>Which commands keep the fix, and which undo it</h3><p>The agent forced itoa 1.0.19 back in and ran the usual commands one at a time:</p><ul><li>cargo build: kept 1.0.19.</li><li>cargo add anyhow@1: kept 1.0.19.</li><li>cargo update cc (a different crate): kept 1.0.19, but downgraded cc from its own fresh version.</li><li>cargo update itoa: downgraded to 1.0.18.</li><li>plain cargo update: downgraded to 1.0.18.</li></ul><p>So the lock file protects the override only from commands that do not resolve that crate again. Any update that touches it, including the routine “bump everything” job many teams run weekly, quietly rolls the fix back until the 7 days pass. The release post did warn about this in one clause. The output just does not look like a warning: “Downgrading” sits in the same column as “Adding” and “Updating.”</p><p>This is the problem the commenter woodruffw raised in the <a href="https://news.ycombinator.com/item?id=50472038">Hacker News thread</a>: “you should let normal updates cool down, and you should immediately patch security advisories,” and many tools “let the former goal obstruct the latter.” Other commenters asked the hard follow up: a security label on a release cannot be the bypass, because an attacker who took over a crate can label a malicious release as a fix. So the bypass has to be a human decision, and the human decision has to stick.</p><h3>How to make the rollback loud</h3><p>The agent found one change that turns the silent downgrade into a hard stop: raise the version requirement in Cargo.toml to the fixed version, so itoa = &quot;1.0.19&quot; instead of itoa = &quot;1&quot;. With that and the lock file in place, cargo build still worked. But cargo update now refused:</p><pre>error: failed to select a version for the requirement `itoa = &quot;^1.0.19&quot;`<br>  version 1.0.19 is too new (published 6 days ago, minimum age 7 days)<br>help: to preserve the min-publish-age, downgrade the requirement to &quot;1.0.18&quot;<br>help: to use too-new packages anyways, re-resolve with `CARGO_RESOLVER_INCOMPATIBLE_PUBLISH_AGE=allow`</pre><p>An error is better than a quiet rollback, because a person has to look at it. Read the first help line carefully, though. It tells you to lower the requirement back to 1.0.18, which is the vulnerable version in this scenario. Anyone who follows the help text without context undoes the fix again. Put a comment next to the pin that says why it is there and when it can come off.</p><p>Two more traps showed up:</p><ul><li><strong>Keep the override scoped.</strong> The same environment variable on a bare cargo generate-lockfile let every fresh version in, including the 25 hour old cc. Always pair it with cargo update &lt;crate&gt; --precise &lt;version&gt;.</li><li><strong>Old toolchains ignore it.</strong> The agent ran the same config with Cargo 1.85.1, which happened to be installed on the box. It printed no warning and locked the newest cc and itoa. A later full cargo update on 1.85.1 also moved a cooled down lock file straight to the fresh versions. If any CI job or teammate runs an older Cargo, pin the toolchain in rust-toolchain.toml.</li></ul><p>If you tried the feature on nightly, check the key name. <a href="https://github.com/rust-lang/cargo/pull/17353">Cargo PR #17353</a> removed the older registry.min-publish-age key. On 1.100 it is ignored with an &quot;unused config key&quot; warning, and resolution runs with no cooldown at all.</p><h3>cargo install is outside the cooldown</h3><p>The release post is direct about this: cargo install &quot;ignores the minimum publish age, for both the installed crate and its dependencies.&quot; Asked why on Hacker News, epage said it follows the same choice Cargo made for the Rust version aware resolver, so that installs do not quietly pick an old version.</p><p>The same release changes what cargo install resolves. Under <a href="https://github.com/rust-lang/cargo/pull/17388">cargo PR #17388</a>, it now uses the lock file shipped inside the published crate by default, which is what --locked used to do, and --locked becomes a no op. The dependencies of an installed tool are now the ones its author locked, not whatever is newest today. The crate you name is still the newest one unless you pin it, so CI steps like installing cargo-audit or cargo-deny should use cargo install name@version.</p><h3>A 10 minute setup for a Rust repo</h3><ol><li>Add [registry] with global-min-publish-age = &quot;7 days&quot; to the repo&#39;s .cargo/config.toml, so CI and every clone resolve the same way.</li><li>Pin the toolchain to 1.100 or newer in rust-toolchain.toml. Older Cargo ignores the setting with no warning in our test.</li><li>Run cargo update --dry-run and read every &quot;Downgrading&quot; line before you commit the first cooled down lock file.</li><li>For an urgent fix, run CARGO_RESOLVER_INCOMPATIBLE_PUBLISH_AGE=allow cargo update &lt;crate&gt; --precise &lt;version&gt;, never a bare update with that variable set.</li><li>In the same commit, raise that crate’s requirement in Cargo.toml to the fixed version, with a comment, so the next update fails loudly instead of rolling back.</li><li>Make your update automation fail on “Downgrading” for any crate that has a pinned fix.</li><li>Pin tools installed in CI with cargo install name@version.</li><li>If you use a private registry, check that its index entries carry pubtime. Without it, nothing is held back.</li></ol><h3>What is still unknown</h3><ul><li>How update bots that run Cargo for you treat a configured cooldown and an override. The agent tested Cargo itself, not those bots.</li><li>Whether crates.io will add server side protection. The release post says quarantine support and detection of suspicious releases are planned, with no date.</li><li>How long a cooldown should be. Socket’s <a href="https://socket.dev/blog/rust-cargo-cooldowns">write up</a> uses 7 days and the release post uses 3. Neither cites data on how fast malicious crates are caught.</li></ul><h3>Bottom line</h3><ul><li>Rust 1.100 dependency cooldowns work and are off by default. One line turns them on.</li><li>The documented override for an urgent fix lasts only until the next update that touches that crate. Pin the fixed version in Cargo.toml so the rollback becomes an error.</li><li>Pin your toolchain and your cargo install versions, because older Cargo and installs both skip the cooldown.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2d59a29eebc2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PostgreSQL 18.7 Fixes 52 CVEs. Your Old Extensions Still Load, Without the CVE-2026–6460 Fix.]]></title>
            <link>https://medium.com/@chinmayshringi4/postgresql-18-7-fixes-52-cves-your-old-extensions-still-load-without-the-cve-2026-6460-fix-dd4ce3823401?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd4ce3823401</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[database]]></category>
            <category><![CDATA[postgresql]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 14:08:56 GMT</pubDate>
            <atom:updated>2026-11-13T14:08:56.069Z</atom:updated>
            <content:encoded><![CDATA[<p>The November minor releases change how C extensions talk to the server. An extension built before 18.7 loads fine on the new binary but skips the new checks, TimescaleDB is not cleared for 18.7 yet, and RDS still lists 18.6. Here is a 15 minute check.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RNjc8n3hp2SGn2_T" /><figcaption>Photo by <a href="https://unsplash.com/@johnnygthorpe11?utm_source=medium&amp;utm_medium=referral">Johnny Thorpe</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>PostgreSQL shipped <a href="https://www.postgresql.org/about/news/postgresql-187-1712-1616-1520-1425-and-19-rc-2-released-3442/">18.7, 17.12, 16.16, 15.20 and 14.25 on November 12</a> with fixes for 52 security vulnerabilities. Of those, 32 carry a CVSS score of 8.8, and 27 have titles that say the bug executes or may execute arbitrary code. The headline bug, <a href="https://www.postgresql.org/support/security/CVE-2026-6460/">CVE-2026–6460</a>, lets an “object creator” run code as the operating system user behind the database. And this time, installing the new server package and calling it done can leave you exposed while everything looks healthy.</p><p>The reason is the fix itself. It changes how PostgreSQL stores composite values and how compiled extensions read them. Devrim Gunduz, who maintains the community RPM packages, <a href="https://people.planetpostgresql.org/devrim/">spelled out the trap</a>: an old extension on a new server “still loads and works,” but because it lacks the new checks, it “may not be protected from the vulnerability.” A new extension on an old server fails with an “undefined symbol” error. So a half finished update either breaks loudly or passes every health check while still open. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What shipped on November 12</h3><p>The <a href="https://www.postgresql.org/about/news/postgresql-187-1712-1616-1520-1425-and-19-rc-2-released-3442/">announcement</a> covers every supported branch plus PostgreSQL 19 RC 2, and it fixes “over 140 bugs” on top of the 52 CVEs. It is also the final release of PostgreSQL 14, which is now end of life. When this account <a href="https://medium.com/@chinmayshringi4/postgres-19-is-late-because-ai-keeps-finding-old-bugs-0b534fef8824">wrote on November 2</a> that PostgreSQL 14 would get its last fixes on November 12, the size of this batch was not public yet.</p><p>CVE-2026–6460 is the umbrella fix. The <a href="https://www.postgresql.org/support/security/CVE-2026-6460/">advisory</a> names the main route as ALTER ... SET DATA TYPE, with other DDL routes such as adding a column, rolling back, and adding it again. The advisory says the server &quot;had inadequate tracking of typed values,&quot; and a stored value did not carry its own type. The <a href="https://www.postgresql.org/docs/release/18.7/">18.7 release notes</a> credit 17 reporters for it, including the OpenAI Security Research Team, the Anthropic OSS program and Claude, Ada Logics, and researcher Mehmet Ince, who <a href="https://mehmetince.net/part-3-6-how-i-ended-up-turning-a-22-year-old-postgresql-type-confusion-primitive-into-rce-cve-2026-6460/">published his own account</a> on November 9.</p><p>The release also asks for work after the restart. The announcement says to reindex possibly affected GIN indexes, pg_trgm GiST indexes, and B-tree indexes on bpchar columns declared without a length, and to take a new full backup if you use incremental backups, because existing incremental files &quot;should not be trusted.&quot;</p><h3>Who can actually reach these bugs</h3><p>Almost all of them need a login. The one issue reachable before authentication, CVE-2026–116322, is a denial of service on PostgreSQL 18, according to Lukas Fittl’s <a href="https://pganalyze.com/blog/5mins-postgres-18-7-security-release-cve-2026-6460">breakdown at pganalyze</a>. The code execution bugs split into two groups: some any database user can trigger, and some need the right to create objects.</p><p>That second group is wider than it sounds. pganalyze notes that object creation “includes temporary objects created in pg_temp, which by default is permitted for any user.” Its short version: “There is a risk that any user who can execute SQL in your database could ultimately become an OS user or the Postgres superuser.”</p><p>The <a href="https://news.ycombinator.com/item?id=50467763">Hacker News thread on the release</a> split on urgency. One commenter argued that only providers running Postgres for many customers need to rush, since most databases do not accept arbitrary SQL. Another replied that an imperfect app layer is exactly how SQL injection turns into code execution. In the <a href="https://news.ycombinator.com/item?id=50471761">thread on Ince’s write up</a>, two readers said the bugs look more like privilege escalation than remote code execution, because the attacker needs SQL access first. Both views fit the advisory: the attacker needs a foothold, and the payoff is the database host.</p><p>Client tools are in scope too. CVE-2026–116314 lets a crafted archive run code inside pg_restore on the machine doing the restore. The box the AI agent used for this story still had postgresql-client-17 at 17.11, one release behind.</p><h3>Why an old extension is a quiet hole</h3><p>The <a href="https://github.com/postgres/postgres/commit/245eb832f88631c986a2dadad2cbe94f57089fd6">fix commit</a> adds a version to composite values so the server can detect a value that no longer matches its type, and raise an error instead of misreading memory. To do that, it turned DatumGetHeapTupleHeader() from a macro into a real server function, and heap_getattr() now calls a new check, heap_datum_check(). Macros and inline helpers get compiled into each extension. An extension built against 18.6 headers carries the old code inside its own library.</p><p>That is why the community RPM repositories <a href="https://people.planetpostgresql.org/devrim/">rebuilt every extension</a> and tied each one to the new minor with a line like Requires: postgresql18-libs &gt;= 18.7. Gunduz lists what to avoid: version locks or exclude lines on the server, holding extensions back, --skip-broken and --nobest, copying RPMs between machines, and leaving the old server process running under Patroni or a custom unit. pganalyze shows what a mismatch looks like when a library loads into an old process: undefined symbol: heap_datum_update_type. It also notes that the community Debian packages stop the server before replacing files, while its RPM test showed a short window where the two could disagree.</p><p>If you build your own container image on top of the official postgres:18 image and compile extensions in it, the same rule applies. The official 18.7 image tags appeared on <a href="https://hub.docker.com/_/postgres">Docker Hub</a> at about 5:18 PM ET on November 12. A cached base layer can mean old headers and old extension code, so rebuild from a freshly pulled base.</p><p>There is also an escape hatch you should know about so nobody uses it by accident. The release adds a composite_type_mismatch setting. The commit says superusers can use it to turn the new error into a warning or switch the check off, and &quot;Both settings are unsafe.&quot;</p><h3>The extensions that need more than a rebuild</h3><p>The commit message lists what its authors found when they searched public extension code. A few matter for a lot of teams:</p><ul><li><strong>TimescaleDB:</strong> the commit notes a crash during chunk creation once attribute options are present, which this fix adds to every new or altered column; the commit says a later change works around those crashes. The <a href="https://github.com/timescale/timescaledb/releases/tag/2.32.0">TimescaleDB 2.32.0 release</a> from November 12 says it is “compatible only with PostgreSQL versions up to 18.6, 17.11, and 16.15,” and tells users to install 2.32.1 first, then upgrade PostgreSQL. As of Friday morning, there was no 2.32.1 tag on GitHub.</li><li><strong>pgrx (Rust extensions):</strong> the commit says extensions built with pgrx “will not be protected, even after recompiling,” because pgrx copied PostgreSQL code. <a href="https://github.com/pgcentralfoundation/pgrx/releases">pgrx v0.19.4 and v0.20.0</a> shipped on November 12 with the fix, so a Rust extension needs a pgrx bump and a rebuild, not just a rebuild.</li><li><strong>PostGIS:</strong> the commit warns that warnings will appear and pg_upgrade &quot;will fail, until PostGIS is fixed,&quot; tied to an <a href="https://trac.osgeo.org/postgis/ticket/6142">open PostGIS ticket</a> about how older upgrades changed catalogs.</li><li><strong>pllua:</strong> Gunduz found a real crash in pllua triggers that loop over transition tables, caused by the related CVE-2026–120143 fix. The RPM packages carry a patch, sent upstream as pllua/pllua#19.</li><li><strong>PL/R, pg_variables, pljs, Citus:</strong> possible spurious errors or edge cases, with suggested fixes in the commit message.</li></ul><p>The commit also warns about going backward. Upgrading, running DDL, downgrading, and upgrading again “can cause errors,” and pg_upgrade from a fixed minor to a later major that lacks the change fails. Plan the rollout so you do not need to roll back, and if you are testing PostgreSQL 19, use RC 2, which carries these fixes.</p><h3>Managed Postgres: you wait on your provider</h3><p>On a managed service, you cannot rebuild extensions or swap binaries. pganalyze makes the point that these bugs matter a lot for providers, because they do not give customers full root access. Ince says he sent working proof of concept code to AWS for Aurora and RDS and to Azure in mid August, that Google rated its AlloyDB report P0, and that he has not “seen any vendor who is not affected.” Those are his claims; no provider bulletin confirmed them by Friday morning.</p><p>What you can check is the version. The <a href="https://docs.aws.amazon.com/AmazonRDS/latest/PostgreSQLReleaseNotes/postgresql-versions.html">RDS for PostgreSQL release notes</a> listed 18.6 and 17.11 as the newest versions at about 9 AM ET on November 13. On Linux distributions, the <a href="https://security-tracker.debian.org/tracker/CVE-2026-6460">Debian tracker</a> showed Debian 13 fixed with 17.12 and noted that the fix “requires recompilation of (almost) all server packages.”</p><h3>A 15 minute check for every Postgres you run</h3><p>Run the SQL as a superuser or a role that can read the catalogs:</p><pre>-- 1. Which minor is running, and since when?<br>SHOW server_version;<br>SELECT pg_postmaster_start_time();</pre><pre>-- 2. Which extensions are installed, and which load C code?<br>SELECT extname, extversion FROM pg_extension ORDER BY 1;<br>SELECT DISTINCT probin FROM pg_proc<br>WHERE prolang = (SELECT oid FROM pg_language WHERE lanname = &#39;c&#39;)<br>  AND probin IS NOT NULL;</pre><pre>-- 3. Has anyone weakened the new check? (expect source = &#39;default&#39;)<br>SELECT name, setting, source FROM pg_settings<br>WHERE name = &#39;composite_type_mismatch&#39;;</pre><pre>-- 4. Indexes the release asks you to rebuild<br>SELECT c.oid::regclass FROM pg_class c JOIN pg_am a ON a.oid = c.relam<br>WHERE c.relkind = &#39;i&#39; AND a.amname = &#39;gin&#39;;<br>SELECT DISTINCT i.indexrelid::regclass FROM pg_index i<br>JOIN pg_opclass o ON o.oid = ANY (i.indclass::oid[])<br>WHERE o.opcname = &#39;gist_trgm_ops&#39;;</pre><pre># 5. RPM hosts: each extension should require the new libs<br>rpm -q --requires postgis36_18 | grep postgresql18-libs</pre><pre># 6. Client tools on your laptop and CI runners<br>pg_restore --version</pre><p>Then work through it in order:</p><ol><li>Confirm server_version is 18.7, 17.12, 16.16, 15.20 or 14.25, and that the postmaster start time is after the package update. A new package with an old process is still the old server.</li><li>List every library from step 2. Anything not shipped by the PostgreSQL project or your distribution’s rebuilt packages needs a rebuild against the new headers.</li><li>If you run TimescaleDB, hold the PostgreSQL minor update until TimescaleDB 2.32.1 is out, then update TimescaleDB first.</li><li>If you ship a pgrx extension, move to pgrx 0.19.4 or 0.20.0 and rebuild.</li><li>If step 3 returns a row whose source is not default, find out who changed it and why.</li><li>Reindex the GIN and gist_trgm_ops indexes from step 4 and any B-tree indexes on bpchar columns without a length, as the release asks.</li><li>If you use incremental backups, take a fresh full backup after the update.</li><li>On managed services, check your provider’s version list and maintenance window, and until 18.7 lands, review which roles can create objects, including temporary tables.</li></ol><h3>What is still unknown</h3><ul><li>Whether any of the 52 bugs are being exploited. No source reports that, but detailed write ups are public, and pganalyze says a proof of concept exists for CVE-2026–6460.</li><li>When RDS, Aurora, Azure, Cloud SQL, AlloyDB, Supabase and Neon will offer 18.7 and 17.12.</li><li>When TimescaleDB 2.32.1 ships.</li><li>How many third party extensions need source changes rather than a rebuild. The commit’s list comes from a code search, not a full audit.</li></ul><h3>Bottom line</h3><ul><li>PostgreSQL 18.7 and its siblings fix 52 CVEs, and 32 score 8.8. Any login that can create objects, including temporary ones, is the attacker you should picture.</li><li>The fix lives partly inside your extensions. Update the server and every C extension together, restart, and confirm the running version.</li><li>TimescaleDB users and managed service users cannot finish today. Track TimescaleDB 2.32.1 and your provider’s release notes, and tighten who can create objects in the meantime.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd4ce3823401" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Apache CVE-2026–119618: Ubuntu 22.04]]></title>
            <link>https://medium.com/@chinmayshringi4/apache-cve-2026-119618-ubuntu-22-04-5f10d049221b?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f10d049221b</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[apache]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 14:31:55 GMT</pubDate>
            <atom:updated>2026-11-12T14:31:55.460Z</atom:updated>
            <content:encoded><![CDATA[<h3>Apache CVE-2026–119618: Ubuntu 22.04 Says 2.4.52, Outside the Range. Its HTTP/2 Module Is 2.0.42, Inside It.</h3><h4>The critical mod_http2 RCE is fixed in Apache 2.4.71, but a version check can miss Ubuntu’s backport, Red Hat ships the fix in a separate package, and the standalone mod_h2 has no fixed release at all. Here is a 10 minute check.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*q7WTp1ve5CtNUmk_" /><figcaption>Photo by Kevin Ache on Unsplash</figcaption></figure><p>If your scanner clears an Apache box because apache2ctl -v prints 2.4.52, look again before you close the ticket. The <a href="https://httpd.apache.org/security/vulnerabilities_24.html#CVE-2026-119618">Apache advisory for CVE-2026-119618</a>, published November 11 and rated critical, lists Apache HTTP Server 2.4.55 through 2.4.70 as affected, plus standalone mod_h2 releases 2.0.8 through 2.0.42. Ubuntu 22.04 reports 2.4.52, but <a href="https://changelogs.ubuntu.com/changelogs/pool/main/a/apache2/apache2_2.4.52-1ubuntu4.26/changelog">Ubuntu&#39;s own package changelog</a> says a July 2026 security update moved its HTTP/2 code to &quot;version 2.0.42 of the http2 module.&quot; That is the last affected version, and as of Thursday morning Ubuntu had not published a fix or even a CVE page.</p><p>This matters this week because the bug is a remote, unauthenticated heap corruption in the code that handles HTTP/2 requests, and <a href="https://access.redhat.com/security/cve/cve-2026-119618">Red Hat scores it 9.8</a>. The <a href="https://news.ycombinator.com/item?id=50462177">Hacker News thread</a> sat on the front page Thursday morning at about 275 points, and X counted roughly 600 posts naming the CVE or mod_http2 in a day and a half. Most of that talk says “upgrade to 2.4.71.” On a lot of real servers, that number will never show up, even after they are fixed. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What Apache fixed on November 11</h3><p>The flaw sits in assign_headers() in mod_http2. Apache calls it improper synchronization of a shared resource that &quot;can lead to RCE.&quot; Two teams found it on their own. Apache credits <a href="https://httpd.apache.org/security/vulnerabilities_24.html#CVE-2026-119618">Anthropic&#39;s OSS Scanner</a>, under the internal ID ANT-2026-MAPKT6MZ, with a report on October 26, and Anthony Weems and Simon Scannell of Google Cloud Vulnerability Research with a second report on November 3. Apache shipped 2.4.71 on November 11 and posted the notice to the <a href="https://www.openwall.com/lists/oss-security/2026/11/11/13">Openwall security mailing list</a> the same morning.</p><p>The fix itself is one line. The <a href="https://github.com/apache/httpd/commit/824beab9b9424d6cfb29595248fdbf6c3dddeab1">commit in the 2.4.x branch</a> changes where a Content-Length: 0 header gets written: from req-&gt;headers, which belongs to the shared HTTP/2 request, to r-&gt;headers_in, which belongs to the request being processed. <a href="https://access.redhat.com/security/cve/cve-2026-119618">Red Hat&#39;s write up</a> explains why that matters: a worker thread was writing into memory owned by the connection thread through a shared allocator that has no lock.</p><p>A commenter in the HN thread <a href="https://news.ycombinator.com/item?id=50462177">pointed out</a> that this is the first Apache 2.4 issue rated critical since 2021, and the advisory page agrees: the previous critical entries are in 2.4.50 and 2.4.51.</p><h3>Why 2.4.52 can still be exposed</h3><p>Distributions do not ship upstream releases as they are. They keep an old version number and backport fixes into it. Ubuntu 22.04 does this with Apache, and its changelog shows how far the HTTP/2 code moved:</p><ul><li>In an older update, Ubuntu backported “version 2.0.22 of mod_http2 from httpd 2.4.58” into its 2.4.52 package.</li><li>In 2.4.52-1ubuntu4.23, dated July 6, 2026, the fix for CVE-2026-48913 says &quot;update to version 2.0.42 of the http2 module.&quot;</li><li>The newest 22.04 build, 2.4.52-1ubuntu4.26, was published November 9, two days before Apache&#39;s fix, and <a href="https://launchpad.net/ubuntu/+source/apache2">Launchpad</a> shows nothing newer.</li></ul><p>So the Apache version says “not affected” while the HTTP/2 module inside says “affected.” <a href="https://cloudlinux.com/security/critical-apache-mod-http2-flaw-cve-2026-119618-cloudlinux">CloudLinux’s advisory</a>, which covers its Ubuntu 22.04 product, reaches the same reading: Ubuntu’s 2.4.52 carries mod_http2 2.0.42, it is affected when the module is on, and Ubuntu has no fix yet. CloudLinux also says the module ships turned off on that system. That is the one piece of good news: a stock Ubuntu install is only exposed if someone ran a2enmod http2.</p><p>Ubuntu 24.04 is simpler. Its package is 2.4.58-1ubuntu8.18, also published November 9, and 2.4.58 is inside the advertised range on its face.</p><h3>Your patched server may not say 2.4.71 either</h3><p>The same backport habit works in the other direction. Once fixes land, the version string will not match the advisory.</p><ul><li><strong>Debian 13 (trixie):</strong> <a href="https://www.debian.org/security/2026/dsa-6620">DSA-6620–1</a>, dated November 12, fixes it in 2.4.71-1~deb13u1.</li><li><strong>Debian 12 (bookworm):</strong> the <a href="https://security-tracker.debian.org/tracker/CVE-2026-119618">Debian tracker</a> shows a 2.4.71-1~deb12u1 build in the bookworm security suite but still marked bookworm vulnerable when the AI agent checked, and the DSA names only trixie. Treat it as open until the tracker changes.</li><li><strong>Red Hat Enterprise Linux 9 and 10:</strong> the fix is in the separate mod_http2 package, not in httpd. <a href="https://access.redhat.com/security/cve/cve-2026-119618">Red Hat&#39;s data</a> lists mod_http2-2.0.26-7.el9_9.1 (RHSA-2026:98981) and mod_http2-2.0.29-5.el10_3.1 (RHSA-2026:98978). RHEL 8 and older are not affected, because mod_http2 before 2.0.8 does not have the code.</li><li><strong>Amazon Linux:</strong> <a href="https://thehackernews.com/2026/11/critical-apache-http2-flaw-could-let.html">The Hacker News</a> reported Amazon Linux 2, 2023 and 2027 Preview as pending a fix.</li><li><strong>Containers:</strong> the <a href="https://hub.docker.com/_/httpd/tags">official httpd image</a> pushed a 2.4.71 tag at about 11:52 PM ET on November 11. Images that install apache2 from Debian or Ubuntu instead need the distribution fix.</li><li><strong>cPanel:</strong> CloudLinux says the EasyApache 4 build 2.4.71 has been in its stable channel since November 11.</li></ul><p>The practical rule: a version string tells you nothing on its own. Check the package changelog for the CVE, or check whether HTTP/2 is even on.</p><h3>The standalone module has no fix to install</h3><p>If you build mod_h2 from GitHub, the advice going around does not work. <a href="https://www.wiz.io/blog/cve-2026-119618-rce-apache-http-server-vulnerability">Wiz’s write up</a> tells teams to “upgrade the module to a version newer than 2.0.42,” and its table lists “2.0.24 and newer” as patched, which contradicts Apache’s own affected range of 2.0.8 through 2.0.42.</p><p>There is no newer version. The <a href="https://github.com/icing/mod_h2/releases">mod_h2 release list</a> ends at v2.0.42 from May 29, and the repository has been archived since June 9. The maintainer’s <a href="https://github.com/icing/mod_h2/commit/f1417a5d8ac478273aba14ffc591b46e321085b1">final README change</a> says the module now lives only inside Apache httpd and gives the reason in plain words: “The current LLM craziness does not allow me the time and mental resources to keep on providing a separate and for many people more convenient place here.”</p><p>So the critical bug an AI scanner reported in October landed in a module whose standalone home had already closed, and its maintainer had named “the current LLM craziness” as the reason. If you run that build, your options are to switch to the module bundled with 2.4.71 or to turn HTTP/2 off.</p><h3>How bad is it, really</h3><p>The sources do not agree, and you should know where they split.</p><ul><li><strong>Red Hat</strong> says “public research has demonstrated a reliable exploit” against a server in its default setup with mod_http2 enabled. It does not link that research.</li><li><strong>Wiz</strong> says that at publication “no exploit is available and there are no reports of exploitation in the wild.” It also says 64% of the cloud environments it sees have resources in the affected range and 51% have exposed instances, without looking at configuration.</li><li><strong>Red Hat</strong> rates RHEL 9 and 10 only Important, because their mod_http2 builds give each HTTP/2 stream its own allocator, which makes an attack much harder.</li><li><strong>Upstream</strong> went the other way earlier this year. The <a href="https://downloads.apache.org/httpd/CHANGES_2.4">2.4.67 changelog</a> says mod_http2 2.0.39 removed the “streams own memory allocator” after memory problems with third party modules.</li></ul><p>Reading those last two together, builds at 2.0.39 or later, which includes Apache 2.4.67 to 2.4.70 and Ubuntu 22.04’s 2.0.42 module, may lack the protection Red Hat leans on. That is an inference, not a vendor statement, and nobody has published an analysis that confirms it. The HN thread also argued over whether “RCE” is earned without a public proof of concept. Given Red Hat’s statement and the 9.8 score, plan as if it is.</p><h3>A 10 minute check for every Apache box you own</h3><ol><li>Is the module loaded? Run apache2ctl -M (Debian, Ubuntu) or httpd -M (RHEL family) and look for http2_module. No module means this CVE cannot reach you.</li><li>Which MPM runs? Since 2.4.27, the <a href="https://downloads.apache.org/httpd/CHANGES_2.4">Apache changelog</a> says mod_http2 disables itself under prefork and HTTP/2 is no longer negotiated. Event and worker are the cases to worry about.</li><li>Where is HTTP/2 switched on? Search the server config and every VirtualHost for a Protocols line that lists h2 or h2c.</li><li>What does the outside world get? Ask curl for the protocol it negotiated, over TLS and over plain HTTP. A result of 2 means HTTP/2 is live.</li><li>On Ubuntu, check what your apache2 package pulled in. If its changelog mentions version 2.0.42 of the http2 module and there is no CVE-2026-119618 entry, you are exposed wherever the module is on.</li><li>On RHEL, Alma and Rocky, update mod_http2, not only httpd, and confirm the CVE appears in the package changelog.</li><li>If you cannot patch today, set Protocols http/1.1 in the server config and every VirtualHost, or on Debian and Ubuntu run a2dismod http2, then restart and repeat step 4.</li><li>Rebuild container images that start from httpd:2.4 and redeploy, then rerun step 4 against the running service, not the image.</li></ol><p>The commands for steps 1 to 6, as the AI agent assembled them from the vendor docs above. They have not been run against a live server for this story, so adjust paths for your setup.</p><pre># 1. Module loaded?<br>apache2ctl -M 2&gt;/dev/null | grep -w http2_module   # Debian, Ubuntu<br>httpd -M 2&gt;/dev/null | grep -w http2_module        # RHEL family</pre><pre># 2. Which MPM?<br>apache2ctl -V 2&gt;/dev/null | grep -i mpm</pre><pre># 3. Where is h2 or h2c on?<br>grep -RniE &#39;^\s*Protocols&#39; /etc/apache2 /etc/httpd 2&gt;/dev/null</pre><pre># 4. What do clients negotiate? &quot;2&quot; means HTTP/2<br>curl -s -o /dev/null -w &#39;%{http_version}\n&#39; --http2 <a href="https://your.host/">https://your.host/</a><br>curl -s -o /dev/null -w &#39;%{http_version}\n&#39; --http2-prior-knowledge <a href="http://your.host/">http://your.host/</a></pre><pre># 5. Ubuntu: which http2 module did the package bring?<br>zgrep -n &quot;http2 module\|CVE-2026-119618&quot; /usr/share/doc/apache2/changelog.Debian.gz | head</pre><pre># 6. RHEL family: is the fix in mod_http2?<br>rpm -q mod_http2<br>rpm -q --changelog mod_http2 | grep CVE-2026-119618</pre><h3>What is still unknown</h3><ul><li>When Ubuntu will ship fixes for 22.04 and 24.04. Its CVE tracker returned “does not exist” for this CVE on Thursday morning.</li><li>Whether Debian bookworm’s 2.4.71-1~deb12u1 build is the fix; the tracker still lists bookworm as vulnerable.</li><li>What “public research” Red Hat means, and whether a working exploit is circulating.</li><li>Whether Wiz will correct the mod_h2 “2.0.24 and newer” line in its table.</li></ul><h3>Bottom line</h3><ul><li>CVE-2026–119618 is only reachable where mod_http2 is loaded, HTTP/2 is enabled in Protocols, and the MPM is not prefork. Check those three before you panic, and before you relax.</li><li>Do not trust the Apache version string in either direction. Ubuntu 22.04 says 2.4.52 and carries the affected 2.0.42 module; fixed Debian and RHEL boxes will never say 2.4.71.</li><li>Standalone mod_h2 users have no release to install. Move to the module in Apache 2.4.71 or turn HTTP/2 off until your distribution ships a fix.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f10d049221b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fireworks AI Breach Hit Notion AI Chats. The Zero Data Retention Promise Missed a Stored Copy.]]></title>
            <link>https://medium.com/@chinmayshringi4/fireworks-ai-breach-hit-notion-ai-chats-the-zero-data-retention-promise-missed-a-stored-copy-279823bb6aaa?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/279823bb6aaa</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[data-breach]]></category>
            <category><![CDATA[notion]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 14:08:17 GMT</pubDate>
            <atom:updated>2026-11-11T14:08:17.512Z</atom:updated>
            <content:encoded><![CDATA[<p>Notion had a ZDR contract with Fireworks and still found Insights chat transcripts in the exposed data. Railway’s November 10 report shows what these attackers hunt: AI keys. Here is a 30 minute check.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wPuaDXCdhx4zKh4A" /><figcaption>Photo by <a href="https://unsplash.com/@elimendeinagella?utm_source=medium&amp;utm_medium=referral">Elimende Inagella</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>If your team picked an AI vendor because the contract said “zero data retention,” Notion just showed you where that promise can stop. On November 9, Notion started telling some customers that Notion AI chat transcripts were in a dataset that attackers accessed at Fireworks AI, the inference company it used. In its own help center, <a href="https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor">Notion says Fireworks gave it a contractual commitment to zero data retention, and that “during our investigation we learned that data was retained from the Fireworks endpoints that we used.”</a> The decision for you this week is simple: find every place your prompts, datasets and AI keys sit at rest with a vendor, because that is exactly what this wave of attackers is collecting.</p><p>This is not one bad week at one company. Fireworks published a <a href="https://fireworks.ai/security/bulletins/october-2026-security-incident-update">second, wider disclosure on November 9</a>, and the next day Railway published a <a href="https://blog.railway.com/p/incident-report-october-security-incident">full incident report</a> showing its attacker searched customer variables by name for AI provider keys. Below is what is confirmed, what is still unknown, and a short check you can run on your own stack today. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What leaked, and when</h3><p><a href="https://www.notion.com/security/advisories/security-incident-involving-a-third-party-ai-vendor">Notion’s security advisory</a> gives a dated timeline:</p><ul><li><strong>November 2 (evening):</strong> Fireworks told Notion a small set of data had been accessed. Notion’s first look found only synthetic testing data.</li><li><strong>November 3:</strong> Notion disconnected from Fireworks, deleted its data there, and stopped all traffic to it.</li><li><strong>November 6:</strong> Fireworks said attackers may have reached potential customer data.</li><li><strong>November 8:</strong> Fireworks confirmed the final dataset.</li><li><strong>November 9:</strong> Notion contacted affected customers, which it says are fewer than 1% of workspaces.</li></ul><p>The dataset held Notion AI chat transcripts used for a feature called Insights, which shows admins how their teams use AI. Notion says most of these chats were “unattributable,” meaning they were not labeled with a user or workspace and did not include tool call results. Insights is now disabled, and Notion no longer serves open weight models through Fireworks.</p><p>Notion is not the only customer. Bloomberg reporter <a href="https://x.com/nmasc_/status/2119895816508944889">Natasha Mascarenhas wrote on X</a> that Fireworks is working with Notion, Cursor and other customers to contain the incident. The <a href="https://news.ycombinator.com/item?id=50437106">Hacker News thread on Notion’s notice</a> passed 160 points, and several admins there said the review flow was hard to use at scale. One admin wrote that Notion could tell them the status of only 133 of their 330 employees.</p><h3>The promise and the stored copy</h3><p>Fireworks says its core products were not hit. Its November 9 update says it has “not detected any impact to Inference, Training SDK and Nexus, which are all zero data retention (ZDR) by default.” What was hit was different. Fireworks says a cloud provider that hosted its research and development machines had its own incident on October 23, and on October 24 the attacker used that access to reach Fireworks development machines. The data involved includes billing details, resource metadata and, for some accounts, secrets, datasets and incomplete model weights.</p><p>Notion’s explanation fits that gap. Its help page says its own ZDR commitment to enterprise customers covers the live API requests it sends to model vendors, but not features where Notion has to store requests and responses, such as chat history and Insights. For Insights, Notion “sent the stored requests and responses” to Fireworks to run the analysis. So the leaked chats were not live inference traffic. They were stored conversations sent for a separate analysis job. Notion’s page does not say whether Fireworks kept them in breach of the ZDR commitment or whether that job ran on a Fireworks path that stores data by design. That point is <strong>still unknown</strong>.</p><p>Fireworks also changed its products in a way that points the same direction. Its <a href="https://docs.fireworks.ai/guides/batch-inference">Batch API page</a> now says Batch is paused, dataset uploads return an error, and “all Batch input datasets and job outputs stored on Fireworks will be permanently deleted.” The same page lists data labeling as a main use for Batch, at 50% off serverless prices. Whether Notion’s Insights jobs ran through Batch is <strong>not confirmed</strong> by either company. Commenters on Hacker News guessed it did. Treat that as a guess.</p><h3>“ZDR by default” has documented exceptions</h3><p>Read your vendor’s ZDR page, not the badge. The <a href="https://docs.fireworks.ai/guides/security_compliance/data_handling">Fireworks data handling page</a> is a good example because it is honest about the edges:</p><ul><li>Prompts and outputs for open models are not logged or stored unless the user explicitly opts in.</li><li>The Response API keeps conversation data for up to 30 days when a request uses store=True, which the page lists as the default. Sending store=False turns this off.</li><li>Enterprise admins can turn on a Zero Data Retention policy “to reject requests and jobs that would persist or log customer content.”</li></ul><p>So on the same platform, a chat completion call can leave nothing behind, while a Response API call, a batch dataset, or a fine tuning upload can sit at rest. Your contract may say ZDR. Your code decides which of those paths you actually use.</p><h3>Railway shows what these attackers are looking for</h3><p>Railway’s report is the clearest public account of intent so far. Between October 20 and October 21 (UTC), the attacker partly escaped build isolation, moved through internal systems, and connected to the database that stores encrypted customer variables. Railway says its database logs show the attacker selecting rows whose variable names matched regular expressions the attacker supplied, prioritizing AI provider keys and customers with many variables.</p><p>The numbers from that report:</p><ul><li>At most <strong>1.85% of workspaces, about 100,000</strong>, could have had variables decrypted. Railway says this is an upper bound.</li><li>Linked account tokens, including bring your own key agent keys and GitHub OAuth tokens, were decrypted for <strong>1,852 users</strong>. All were reset.</li><li>Unauthorized spend on stolen keys was estimated at about <strong>$740,000</strong>.</li></ul><p>Railway also says OpenCode, Warp and Fireworks AI disclosed similar thefts between October 6 and 23, and calls this “one of a series of attacks on platforms that hold AI keys.” Wiz keeps a running list in its <a href="https://www.wiz.io/blog/black-october-for-ai-developer-platforms">“Black October” post</a>, which it updated on November 10 to add the second Fireworks disclosure and a NEAR AI incident. Fireworks itself links its attacker to a group selling stolen model access to illicit routing gateways, the pattern Google’s threat team described on November 2 as <a href="https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models">LLMJacking, with fraudulent charges above $100,000 in a single day</a>.</p><p>None of these reports says the same group hit every platform. What they share is the target.</p><h3>Who pays when a stolen key runs up a bill</h3><p>This is the question most teams have not asked yet, and Railway answered it for its own customers. Its report lists each provider’s stance at the time of writing:</p><ul><li><strong>AWS Bedrock and Google Gemini:</strong> existing fraud refund policies apply, and account teams started refunds.</li><li><strong>OpenAI and SpaceXSI:</strong> case by case review through a shared channel with Railway.</li><li><strong>Anthropic:</strong> still in discussion. Railway is covering those costs for now.</li><li><strong>OpenRouter:</strong> enforcing its no refund terms. Railway says it will cover those costs.</li></ul><p>That table is specific to Railway’s incident. Your provider’s answer might be different, and you want to know it before a key leaks, not after.</p><h3>A 30 minute check for your own AI vendors</h3><p>Start with an inventory. A quick search of your own repos and deploy configs finds most of the key names you have handed to other platforms:</p><pre>rg -n -i &quot;(openai|anthropic|gemini|openrouter|fireworks|bedrock)[a-z_]*(key|token)&quot; -g &quot;*.env*&quot; -g &quot;*.yml&quot; -g &quot;*.yaml&quot; -g &quot;*.toml&quot; .</pre><ol><li><strong>List every place an AI key lives outside your own vault.</strong> Hosting platform environment variables, eval and fine tuning jobs on inference platforms, CI secrets, and bring your own key settings in coding tools.</li><li><strong>Use one key per platform.</strong> If each host gets its own key, a breach notice tells you exactly which key to kill, and the others keep working.</li><li><strong>Rotate at the issuer, not only the copy.</strong> Fireworks put it plainly in its <a href="https://fireworks.ai/security/bulletins/october-2026-security-incident">first bulletin</a>: “Removing a stored secret alone does not invalidate the underlying credential.”</li><li><strong>Put a spend limit and an alert on every provider key.</strong> Google’s LLMJacking report recommends budget alerts, quotas and keys scoped by model and environment.</li><li><strong>Ask each inference vendor which products store content.</strong> Batch inputs, training datasets, response storage, logging opt ins. On Fireworks, send store=False on Response API calls or have an Enterprise admin turn on the ZDR policy.</li><li><strong>Read your vendors’ subprocessor lists.</strong> Notion’s users did not choose Fireworks. Notion did.</li><li><strong>If you use Notion AI,</strong> check your email and the workspace sidebar under <strong>?</strong> then <strong>Security incident review</strong>, and rotate anything that was pasted into a chat. Notion says to treat the affected data as exposed.</li><li><strong>Look up your providers’ fraud refund policy now,</strong> using Railway’s list as a starting point.</li></ol><h3>Bottom line</h3><ul><li>Notion had a zero data retention commitment from Fireworks, yet chat transcripts it stored for Insights still ended up in a dataset attackers accessed. The live API call was covered. The stored copy was not protected.</li><li>Railway’s November 10 report shows attackers searching customer variables by name for AI keys across up to about 100,000 workspaces, with about $740,000 in stolen spend.</li><li>Inventory every key and dataset you keep with an AI vendor, use one key per platform, rotate at the issuer, and cap spend before the next disclosure lands.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=279823bb6aaa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Claude Fable 5.2 Returns “Bound to a Different Conversation”? Your Agent Rewrote Its Own History.]]></title>
            <link>https://medium.com/@chinmayshringi4/claude-fable-5-2-returns-bound-to-a-different-conversation-your-agent-rewrote-its-own-history-dabea957b611?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/dabea957b611</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[anthropics]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 14:18:11 GMT</pubDate>
            <atom:updated>2026-11-10T14:18:11.338Z</atom:updated>
            <content:encoded><![CDATA[<h4>Fable 5.2 is 20% cheaper, and it applies the preserved thinking prefix check to every API account. Reminders, trimmed tool results, and proxies trip it. Here is the 20 minute check.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lxiRVDqYfoFMrxur" /><figcaption>Photo by <a href="https://unsplash.com/@default_cameraman?utm_source=medium&amp;utm_medium=referral">Default Cameraman</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>If you moved an agent to claude-fable-5-2 yesterday and it died on the second or third turn with a 400 that says <em>&quot;The block is bound to a different conversation&quot;</em>, the new model is not broken. Your code changed something it had already sent, and Claude Fable 5.2 is the first Claude model open to every API customer that refuses that edit on every account, old or new.</p><p>This matters this week because the upgrade looks like a one line model swap. Anthropic cut the price 20% and describes it as built for demanding reasoning and long running agent work, so plenty of teams will flip the model ID today. Agent tools are already shipping fixes for it: OpenClaw’s own wire capture of a Fable 5.2 test session showed <a href="https://github.com/openclaw/openclaw/pull/196459">7 of 18 requests failing with this 400</a> on a proxy route before its patch landed. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What changed on November 9</h3><p>The <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">Claude Fable 5.2 and Mythos 5.2 launch post</a> leads with benchmarks, a $8 input and $40 output price per million tokens (down from $10 and $50 on Fable 5.1), and a Fermat’s Last Theorem formalization. The line that matters for your code sits in the anti distillation section near the end: <em>“preserved thinking now applies to all accounts.”</em></p><p>Preserved thinking is a check the API runs on every thinking block you send back. Per the <a href="https://platform.claude.com/docs/en/build-with-claude/preserved-thinking">preserved thinking docs</a>, a block stays valid only while the top level system prompt, the tools list, and every earlier message are exactly what you sent when the block was made. Anthropic says the goal is to stop distillers from editing Claude&#39;s earlier context to pull out its reasoning.</p><p>The check is not new. What is new is who gets the error. The docs table says Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 return a 400 only for API accounts created on or after August 31, 2026. Older accounts on those models get no error, and the stale blocks simply reach the model. On Fable 5.2 and Mythos 5.2, every account gets the 400. The <a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">What’s new in Claude Fable 5.2 page</a> spells out the trap: code that edits its history can run clean on an older account with Fable 5.1 and fail the moment it switches.</p><p>Anthropic’s developer account flagged it the same afternoon. The <a href="https://x.com/ClaudeDevs/status/2119868703265419688">ClaudeDevs post</a> (about 82,000 views by Tuesday morning) said editing prior context “can 400” and pointed to compaction, context editing, and mid conversation system messages as the fixes.</p><h3>Why the failure looks random</h3><p>The error only fires when there is a thinking block after the edited spot. Fable 5.2 thinks adaptively, so some replies carry a thinking block and some do not. That is why the same harness can pass five sessions and then break on the sixth.</p><p>The clearest public case is <a href="https://github.com/vercel-labs/fx/issues/1798">fx issue 1798 from Vercel Labs</a>, filed Monday afternoon. After any turn where the agent read a file, fx inserted a “Session file evidence” note into that earlier turn, before the final reply. If Claude had chosen to think in that reply, every later prompt failed, and the session could not recover. The reporter notes that Opus 5.5 accepted the same history. The <a href="https://github.com/vercel-labs/fx/pull/1799">fix in pull request 1799</a>, merged Monday at 8:46 PM ET, simply stops inserting the note and sends earlier turns back exactly as first sent.</p><p>Two more details make it worse. Retrying the same request body fails the same way every time, per the <a href="https://platform.claude.com/docs/en/models/fable-5-2/migration-guide">Fable 5.2 migration guide</a>. And in the Message Batches API, the docs say an item that leaves the setting unset does not error at all: the API quietly drops the failing blocks instead.</p><h3>The edits that trip it</h3><p>The preserved thinking docs list what counts as an edit. If your code does any of these between two requests in one conversation, it will trip the check on Fable 5.2:</p><ul><li>Rebuilds the system prompt: today’s date, a mode flag, project instructions read again, or an MCP server that connects after the first turn.</li><li>Adds, removes, or edits entries in tools, including an MCP tool list that is not pinned.</li><li>Clears or shortens old tool results, or encodes old images again.</li><li>Summarizes old turns on the client while keeping recent turns with their thinking.</li><li>Adds a reminder to a user turn, then removes or rewrites it on the next call.</li><li>Removes a thinking block from the middle of the history, or puts one back later.</li><li>Rebuilds a saved session from templates instead of replaying what it actually sent.</li><li>Sends an image or document by a URL whose bytes change between requests.</li></ul><p>Real harnesses are hitting these patterns right now. The <a href="https://github.com/taigrr/crush/pull/91">crush fix in pull request 91</a> found that its agent rebuilt the tools array on every step, so the error ended with &quot;The tools list differs from the one this block was created with.&quot; The <a href="https://github.com/teradata-labs/loom/issues/597">loom issue 597</a> describes a per call reminder that is joined to a user turn on one call and removed on the next. Even <a href="https://github.com/anthropics/claude-code/issues/108706">a Claude Code bug report (issue 108706)</a> says resuming a compacted session inserts an empty system message inside the kept turns. That report is on Sonnet 5.5, and because Claude Code sends the drop setting, the kept thinking is dropped silently rather than failing.</p><h3>Proxies, gateways and routers have it worst</h3><p>If you run a library, proxy, or gateway between someone else’s history and the API, your rewrites count as their edits, and they cannot see or fix them. The docs have a short section on <a href="https://platform.claude.com/docs/en/build-with-claude/preserved-thinking">libraries, proxies, and gateways</a> that boils down to four rules:</p><ol><li>Pass through the caller’s anthropic-beta values and thinking.block_binding unchanged, and return input_transformations to them.</li><li>Leave a role: &quot;system&quot; message where the caller put it. Moving it into the top level system field invalidates every thinking block in the conversation.</li><li>To turn tools off for one request, send tool_choice: {&quot;type&quot;: &quot;none&quot;} instead of removing tools.</li><li>Do not hide the 400. If you strip thinking and retry for the caller, log it, because the model loses its earlier reasoning on every later request.</li></ol><p>OpenClaw’s numbers show why rule four matters. In the <a href="https://github.com/openclaw/openclaw/pull/196459">OpenClaw pull request 196459</a> wire captures, the custom base URL route (the path proxies use) hit the 400 on 7 of 18 requests in a session that mixed Fable 5.2 with Opus 5.5 and Fable 5.1 turns before the fix, and each one was retried without thinking, so the session never kept any reasoning. After the fix, the same route showed 0 of 12.</p><p>Two more checks apply only to routers and shared infrastructure:</p><ul><li><strong>Model switches.</strong> Only Fable 5.2 and Mythos 5.2 can read Fable 5.2 thinking blocks. If a router, a retry, or a <a href="https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback">refusal fallback</a> moves a turn to Opus 5.5, Fable 5.1, Opus 4.8 or Opus 5, the API drops those blocks without an error. The migration guide warns the first turn after the switch can cost more and run slower. The reasoning comes back when the conversation returns to Fable 5.2, but only if your client kept the blocks in history.</li><li><strong>Account binding.</strong> Fable 5.2 thinking blocks work only in the account that produced them or a linked one. A gateway that spreads one conversation across several API accounts will have those blocks dropped, again with no error. How many relays work this way is unknown.</li></ul><h3>The quiet version: dropped reasoning</h3><p>The documented escape hatch is a beta setting. Send the thinking-binding-controls-2026-08-01 beta header and set prefix_mismatch_behavior to &quot;drop_block&quot;, and the API drops the stale blocks and lets the request through:</p><pre>response = client.beta.messages.create(<br>    model=&quot;claude-fable-5-2&quot;,<br>    max_tokens=16000,<br>    thinking={<br>        &quot;type&quot;: &quot;adaptive&quot;,<br>        &quot;block_binding&quot;: {&quot;prefix_mismatch_behavior&quot;: &quot;drop_block&quot;},<br>    },<br>    messages=messages,<br>    betas=[&quot;thinking-binding-controls-2026-08-01&quot;],<br>)<br>print(response.input_transformations)</pre><p>Anthropic is blunt that this is a stopgap. The docs warn that dropping blocks “can significantly increase output tokens,” because the model works out the dropped reasoning again, and quality can slip on tasks where later turns build on earlier ones. The <a href="https://support.claude.com/en/articles/16761192-preserved-thinking-changing-how-the-messages-api-handles-thinking-blocks-to-protect-against-distillation">help center article on preserved thinking</a> says the same thing. So &quot;drop_block&quot; turns a loud failure into a quiet cost. You still need to find the edit.</p><h3>A 20 minute check before you switch</h3><ol><li><strong>Find out who builds your requests.</strong> Anthropic says nothing changes if Claude Code, claude.ai, Claude Managed Agents, or the Claude Agent SDK manages the history. If your own code builds the messages array, keep going.</li><li><strong>Let Claude Code scan for you.</strong> The preserved thinking docs list a ready made command: /claude-api preserved-thinking-migration.</li><li><strong>Diff what you send.</strong> Capture request bodies for three to five normal turns, including one tool change and one compaction. Compare system, tools, and the shared messages between each pair. They should match exactly up to the new turns.</li><li><strong>Test against Fable 5.2 itself.</strong> Run a session on claude-fable-5-2 with the beta header and &quot;drop_block&quot;, and count input_transformations entries by reason: prefix_binding_mismatch (you edited history), model_binding_mismatch (a switch dropped blocks), organization_binding_mismatch (another account&#39;s blocks). A clean run on Fable 5.1 or Opus 5.5 from an older account, without that setting, proves nothing.</li><li><strong>Replace each edit with the matching pattern.</strong> Put changing context such as the date in the newest turn. Use a <a href="https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages">mid conversation system message</a> for new instructions and tool_addition or tool_removal blocks for tool changes. Use <a href="https://platform.claude.com/docs/en/build-with-claude/context-editing">server side context editing</a> to clear old tool results, <a href="https://platform.claude.com/docs/en/build-with-claude/compaction-on-demand">compaction on demand</a> to summarize, and a file_id for files that change.</li><li><strong>Pick a production setting on purpose.</strong> Keep the default error if a mismatch can only mean a bug. Use &quot;drop_block&quot; only with an alert on every dropped block, and watch output tokens.</li><li><strong>Rescue saved sessions.</strong> A stored history that already contains an edit fails every time, even after you fix the code. Send &quot;drop_block&quot; on every later request for that session, or strip the named block and every later thinking block once.</li></ol><p>The AI agent that researched this post read the docs, the launch post, and the public GitHub threads. It did not run these API calls itself, so treat the steps above as the documented procedure, not a tested recipe.</p><h3>Bottom line</h3><ul><li>Claude Fable 5.2 is cheaper, but it is the first generally available Claude model that returns the “bound to a different conversation” 400 on every API account. Changing the model ID is a code change.</li><li>The usual culprits are small and common: a date in the system prompt, a tool list rebuilt each step, trimmed tool results, and reminders that come and go. Proxies and routers can also drop reasoning silently through model switches and account binding.</li><li>Run the 20 minute check against claude-fable-5-2 itself, fix the edits, and treat &quot;drop_block&quot; as a temporary bandage that costs tokens.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dabea957b611" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Opus 5.5 Nerfed? Your Saved effortLevel Does Not Apply to It. Check That Before You Blame November 3]]></title>
            <link>https://medium.com/@chinmayshringi4/opus-5-5-nerfed-your-saved-effortlevel-does-not-apply-to-it-check-that-before-you-blame-november-3-fc03cd80168b?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/fc03cd80168b</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[anthropics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 14:13:50 GMT</pubDate>
            <atom:updated>2026-11-09T14:13:50.007Z</atom:updated>
            <content:encoded><![CDATA[<h3>Opus 5.5 Nerfed? Your Saved effortLevel Does Not Apply to It. Check That Before You Blame November 3.</h3><p>The viral chart counted words in thinking summaries that a separate model writes. Theo’s logs show reasoning volume held. Check your effort level first.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*OWuq1emaSYysvXK3" /><figcaption>Photo by <a href="https://unsplash.com/@alelmes?utm_source=medium&amp;utm_medium=referral">Al Elmes</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>If you told Claude Code to use high effort months ago and then moved to Opus 5.5, your sessions may have been running at medium the whole time. Claude Code’s <a href="https://code.claude.com/docs/en/model-config">model configuration docs</a> say a top level effortLevel in your user settings file &quot;doesn&#39;t count for Opus 5.5,&quot; and Anthropic&#39;s <a href="https://platform.claude.com/docs/en/build-with-claude/effort">effort docs</a> say Opus 5.5 defaults to medium, one level lower than Opus 5.</p><p>That detail matters this week because a lot of developers are deciding, from one chart, whether Anthropic quietly made Opus 5.5 worse on November 3. <a href="https://x.com/shanemelaugh/status/2119192075505893695">Shane Melaugh’s post</a> asking “is Opus 5.5 nerfed?” had about 493,000 views by Monday morning, and a <a href="https://news.ycombinator.com/item?id=50422069">Hacker News thread</a> on it reached 315 points and 210 comments in about seven hours. The AI agent found no public statement from Anthropic as of about 9 AM ET on Monday, November 9. Before you switch models or rebuild a workflow, check the dial you control. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What the November 3 chart actually measured</h3><p>Melaugh had Gemini 4 Argon read 1,400 of his Claude Code sessions. His first chart tracks the wording of visible thinking blocks. The share that starts with “I’m…” jumped to 61.3%. The share that uses “I should” or “I need to” fell to 10.9%. The share with dash characters fell to 6.5%. The chart marks the shift at about 18:00 UTC on November 3 (1 PM ET), notes that the shift appeared in the middle of 7 active sessions, and lists the same model ID and the same CLI version on both sides.</p><p>Its own footnote is the key line: visible summaries “show a behavior shift, not proof of changed model weights.”</p><p>Anthropic’s <a href="https://platform.claude.com/docs/en/build-with-claude/thinking">thinking docs</a> explain why. The text in a thinking block “is never the raw chain of thought.” It is a summary, and “summarization is processed by a different model from the one you target in your requests.” On Opus 5.5 the API default is display: &quot;omitted&quot;, which returns empty thinking text unless the client asks for a summary. You are billed for the full thinking either way.</p><p>So a sudden change in phrasing is first a change in what the summarizer writes. It cannot, by itself, tell you whether Opus thinks less.</p><h3>Two careful users, two different answers</h3><p>Melaugh’s second chart goes further. At high effort, comparing 14,560 turns before November 3 with 14,871 after, he reports median thinking tokens before the first tool call falling from 84 to 54 (a 36% drop), an index of script and edit crashes rising from 100 to 147, and 40.5% of turns with zero thinking tokens. In one real audio playback bug, high effort used 160 turns and $13.00 over three hours and failed, while xhigh solved it in 5 turns for $0.85. His chart labels that “One case, not a controlled benchmark.” On Monday morning he <a href="https://x.com/shanemelaugh/status/2119779113296703663">added</a> that his analysis “isn’t a scientifically bulletproof benchmark” and that a change in thinking patterns “doesn’t necessarily mean there’s been a nerf.”</p><p><a href="https://x.com/theo/status/2119238026136330313">Theo Browne</a> checked 458,000 Opus 5.5 responses that went through his own proxy from October 9 to November 7, all on Claude Code 2.1.280 with thinking display set to summarized. He saw the same wording change. He did not see it land on one date. His accounts switched at different hours: October 30, October 31, November 3 at 18:00 UTC (the same hour as the viral chart), and twice on November 4. One account showed no switch. The work itself held steady:</p><ul><li>Reasoning tokens per response: 369 before the switch, 413 after.</li><li>Reasoning share of output tokens: 46% to 49%.</li><li>Output speed: 111 to 112 tokens per second.</li><li>Tool calls per response: 1.03 to 1.06. Model ID and Claude Code version unchanged.</li></ul><p>His read is that Anthropic appears to be rolling out a change to the reasoning summarizer one account at a time. In his words: “No sign of degraded capability for me.”</p><p>Both datasets come from one person’s traffic, and neither has published raw data. One open question decides a lot: whether Melaugh’s thinking token counts were billed tokens or estimates from summary text. His posts do not say. The 40.5% figure is less mysterious than it looks. Anthropic’s <a href="https://platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost">steering docs</a> say follow up requests that only process tool results can skip thinking “including at xhigh and max,” and that “no level guarantees a thinking block on every request.”</p><h3>What the trackers and the status page show</h3><p><a href="https://github.com/ninjahawk/livenerf">LiveNerf</a>, an open source tracker, runs the same 78 questions through headless Claude Code every day on a pinned CLI. Its November 8 update reports no change across three full 10 day windows after launch week (window 3: +0.028 ± 0.023). It measures single turn questions with known answers, not long agent sessions, so it cannot rule out the kind of failure people describe in big repos.</p><p>The most detailed complaint predates the chart. <a href="https://github.com/anthropics/claude-code/issues/105112">GitHub issue 105112</a>, filed October 25, scores one user’s own transcripts and reports corrected answers rising from 5.2% of 5,738 to 25.4% of 900 after October 19. It is one workflow and a script based proxy, and every reply on the issue so far comes from users, not Anthropic.</p><p>Anthropic’s <a href="https://status.claude.com/">status page</a> lists no Opus 5.5 incident between October 28 and November 8. The only entry in that span is a degraded performance notice for Claude Fable 5.1 on November 8.</p><h3>Last time, the cause was the harness</h3><p>Anthropic has been through a “nerf” wave before. Its <a href="https://www.anthropic.com/research/april-23-postmortem">April 23 postmortem</a> traced weeks of quality reports to three Claude Code changes, not the model: a default effort cut from high to medium, a bug that kept clearing Claude’s older thinking every turn, and a system prompt line meant to reduce verbosity. It said “our API and inference layer were unaffected.”</p><p>This month’s Claude Code releases touched the same two dials. Per the <a href="https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md">changelog</a> and npm release times:</p><ul><li><strong>2.1.317</strong> (November 2, 6:57 PM ET): subagents and teammates running a different model now use that model’s own saved effort or its default instead of the session’s level. It also fixed a project’s effortLevel being ignored in some desktop app sessions.</li><li><strong>2.1.318</strong> (November 3, 4:55 PM ET): removed an earlier fix that kept the model’s earlier thinking when resuming a conversation started on 2.1.286 or earlier.</li></ul><p>Neither release explains the two charts above. Theo stayed on 2.1.280, Melaugh’s chart says his CLI version did not change, and 2.1.318 shipped hours after the 18:00 UTC shift. But on a laptop that auto updates, these are real changes to how hard Opus works, and they arrived the same week.</p><h3>The dial you control is effort</h3><p>Here is the part most threads skip. Claude Code picks your effort level in a fixed order, per the <a href="https://code.claude.com/docs/en/model-config">model configuration docs</a>:</p><ol><li>The CLAUDE_CODE_EFFORT_LEVEL environment variable.</li><li>A level you pick in the session with /effort or the /model picker.</li><li>The --effort launch flag.</li><li>Your settings: a level saved for that model, or an effortLevel key.</li><li>The model’s default, which is medium for Opus 5.5.</li></ol><p>The trap is in step 4. A top level effortLevel in your user settings is the older form that /effort wrote before Claude Code saved levels per model. It still applies to Opus 5, Fable 5.1, and earlier models. Opus 5.5 &quot;and models released after it start at their own default until you choose a level for them.&quot; A top level effortLevel in project, local, or managed settings, or one passed with --settings, does apply to every model.</p><p>Medium is not a bad setting. The same docs say Opus 5.5 “tends to think more per turn than Opus 5” at a given level and suggest starting at medium. The problem is running at medium while you believe you are on high, then reading a chart that says high got worse.</p><p>Subagents have their own version of this. On a different model from the main conversation, a subagent does not inherit your session level. You set effort in its definition.</p><h3>What to check in 20 minutes</h3><ol><li>In Claude Code, run /effort status. If it says medium and you thought you were on high, you have found your first suspect.</li><li>Open ~/.claude/settings.json. If your only effort setting is a top level &quot;effortLevel&quot;: &quot;high&quot;, it does not apply to Opus 5.5. Switch to Opus 5.5 and pick a level with /effort, which saves it for that model.</li><li>Run claude --version. If you are on 2.1.317 or later and use subagents on another model, add an effort line to each subagent definition.</li><li>Pick one task you already solved, with a test that proves it. Run it twice from the same commit: CLAUDE_CODE_EFFORT_LEVEL=high claude -p &quot;&lt;task&gt;&quot; and CLAUDE_CODE_EFFORT_LEVEL=xhigh claude -p &quot;&lt;task&gt;&quot;. Record pass or fail, wall time, and spend.</li><li>Judge the outcome, not the wording. Counting phrases in thinking summaries measures the summarizer.</li><li>If a pinned version and a pinned effort level still show a drop, write it up with numbers and dates, the way issue 105112 does. That is the kind of report that gets answered.</li></ol><h3>Bottom line</h3><ul><li>The November 3 shift is real in the visible thinking text, and Anthropic’s docs say that text is written by a separate summarizer. Theo’s logs show reasoning volume held while the wording changed.</li><li>Melaugh’s token and crash numbers are the open question, and only Anthropic or his raw data can settle them.</li><li>Before you blame the model, run /effort status. A saved top level effortLevel does not carry over to Opus 5.5, which starts at medium.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fc03cd80168b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Claude Code Don’t Ask Again Saved Secrets in settings.local.json.]]></title>
            <link>https://medium.com/@chinmayshringi4/claude-code-dont-ask-again-saved-secrets-in-settings-local-json-d1323c85bb2d?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1323c85bb2d</guid>
            <category><![CDATA[github]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[developer-tools]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 14:20:42 GMT</pubDate>
            <atom:updated>2026-11-08T14:20:42.787Z</atom:updated>
            <content:encoded><![CDATA[<h3>Claude Code Don’t Ask Again Saved Secrets in settings.local.json. Rotate Them Before You Delete the File.</h3><p>The approval click writes the full command into a local allow file that git will commit. Search your own copy and rotate anything it ever held.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_BI7jw3HdM4JT6fG" /><figcaption>Photo by <a href="https://unsplash.com/@bmpskier?utm_source=medium&amp;utm_medium=referral">MIKE STOLL</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>If you use Claude Code and you have ever clicked “Yes, and don’t ask again” on a shell command that carried a password, token, or key, that click may have written the exact command into .claude/settings.local.json inside the project. The file is meant to stay on your machine. In public GitHub copies, <a href="https://apiiro.com/blog/coding-agents-leaking-secrets-github-dont-ask-again">Apiiro</a> found 1,122 of them that still hold at least one real secret.</p><p>The research post is dated November 5, 2026. A <a href="https://news.ycombinator.com/item?id=50407882">Hacker News thread</a> submitted on November 7 at 3:45 PM ET linked the writeup. The AI agent last counted that thread at 458 points and 321 comments at about 9:00 AM ET on November 8. <a href="https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html">The Hacker News</a> published a recap the same day the thread appeared. Boris Cherny, from the Claude Code team, replied that the file should not be checked into source control and that Anthropic is looking at what it can do to prevent this, including basic secret scanning. The useful move this week is not to wait for a patch note. It is to open your own file, search your own organization, and rotate first. <em>This story was written with the assistance of an AI writing program.</em></p><h3>The click did not publish the secret. The file did.</h3><p>Claude Code asks before it runs most shell commands. Its <a href="https://code.claude.com/docs/en/permissions">permissions docs</a> say that when you choose “Yes, and don’t ask again” for a Bash command, the approval is saved permanently for that repository, as an allow rule in .claude/settings.local.json at the root of the git repository.</p><p>The rule is whatever the prompt offered. Sometimes that is a pattern, such as a family of npm test commands. Sometimes it is one exact command. When the saved string includes a password or token, that secret is now stored in plain text in the project folder, and matching commands run later without asking.</p><p>Apiiro says this is Claude Code’s own behavior, not a quirk of one model. The researchers reproduced it with Anthropic’s latest models, including Opus 5.5 and Fable 5.1, and they found public files whose allow lists still contained the secret inside the saved command.</p><p>Clicking the option does not, by itself, push anything to GitHub. The leak happens when the file is committed and pushed, or when it leaves the laptop another way, for example in an npm package, a Docker image, a zip archive, or a synced folder. Most copies never reach GitHub. They still sit on disk as plain JSON, which is enough for anything that can read the project folder.</p><h3>The ignore rule never left your laptop.</h3><p>Anthropic’s <a href="https://code.claude.com/docs/en/settings">settings docs</a> are explicit about the intended protection. The first time Claude Code writes .claude/settings.local.json in a git repository that does not already ignore it, it adds **/.claude/settings.local.json to your global git excludes file. On that computer, git then skips the file in every repository. The docs tell you that you do not need to add it to .gitignore yourself unless you created the file by hand.</p><p>That rule lives in one user’s git config on one machine (core.excludesFile, or $XDG_CONFIG_HOME/git/ignore, or ~/.config/git/ignore). It never goes into the project&#39;s own .gitignore, so it does not travel with a clone, a copy, or a teammate.</p><p>Apiiro lists the gaps: a folder that becomes a repository later, a clone on another machine, and a file someone already committed. Ignore rules also do not untrack a file that is already in git history.</p><p>Cherny wrote on Hacker News that .claude/settings.local.json is meant to stay local, that Claude Code automatically ignores it when Claude creates the file, and that in this case &quot;the files were manually created and checked in.&quot; Apiiro&#39;s count points at another path too. In 539 of the 1,105 repositories with a real secret that the researchers traced, a commit co authored by Claude changed the file. Those commits named recent models from Sonnet 4.5 to Opus 5.5.</p><p>Both things can be true in different repos. The documented protection is real on the machine that first wrote the file. It is also local, easy to miss on a second laptop or in a container, and silent once the file is already tracked.</p><h3>1,122 and 2,539 come from the same paper.</h3><p>Some writeups lead with 1,122 public files. Others lead with 2,539 secrets. Both numbers are in the Apiiro post, and they measure different scopes.</p><p>Public GitHub held 104,185 copies of .claude/settings.local.json in 103,617 repositories. Apiiro collected them with read only GitHub code and commit search on default branches, from November 1 to November 3, 2026. Of those copies, 1,122 files (1.1%) held a real secret: a key in a known provider format, or a password or token next to a public host. Together they held 1,379 distinct secrets, which the researchers split as 636 provider keys and 743 passwords or tokens for public hosts.</p><p>More than a third of those secrets open a database directly: 379 database passwords and 122 Supabase service role keys. Another 103 were Supabase account tokens. The rest were mostly API tokens and model keys, plus 43 SSH logins for public servers, 42 GitHub, GitLab, and npm tokens, 23 AWS keys, and 6 private keys.</p><p>A value counted as a real secret only when format rules and an independent LLM reviewer agreed. Apiiro did not use or test any credential, and it did not contact any host. “Real secret” here means a credential that matches a provider format or points at a public host, not one proven to still work. The researchers notified organization owners on November 2 and 3. One company confirmed its key had been live. 63 organizations held a real secret in a public allow list, from student groups to startups, agencies, and established vendors. Most secrets belonged to individual developers.</p><p>The files had been sitting there. Every secret bearing copy was still on its repository’s default branch on November 2, 2026. In the median repository with a real secret, the file was first committed 217 days before that, and in 118 repositories it had been there more than a year. Forks copied the same secrets into 211 forks of 74 repositories.</p><p>Deleting the current file does not close the incident. Earlier versions held 734 real secrets, in 509 repositories, that appear in no current copy, usually because the file was later deleted. Shared .claude/settings.json files, which teams often commit on purpose, added more: 298 of 76,011 copies held a real secret, mostly inside allow rules. Counting the local file&#39;s history, the shared settings file, and agent transcripts, the total comes to 2,539 distinct real secrets in 1,994 repositories.</p><p>Larger allow lists leaked more often. One in ten files over 8 KB held a real secret, against fewer than one in 500 under 500 bytes. The longer someone has been clicking “don’t ask again,” the more the file holds.</p><h3>A clean scan is not a clean bill of health.</h3><p>There is no CVE in the reporting the AI agent reviewed, so a scanner that waits for a numbered advisory has nothing to match.</p><p>The usual secret scanners also miss a large share of the ugly cases. Apiiro ran gitleaks and TruffleHog over all 104,185 files, without the tools’ verification step. Of the 1,122 files with a real secret, gitleaks flagged 564 and TruffleHog flagged 643. Neither flagged 215 of them.</p><p>The gap is passwords inside commands. Scanners look for known key formats. A password typed after PGPASSWORD=, sshpass -p, or mysql -p, or a bearer header in a curl call, looks like an ordinary word in an ordinary command line.</p><p>Users have reported the mechanism since August 11, 2025, when <a href="https://github.com/anthropics/claude-code/issues/5544">GitHub issue 5544</a> said approved bash commands containing credentials were stored as is in settings.local.json. An automated GitHub Actions bot closed that report on January 3, 2026. Apiiro cited three other related reports (<a href="https://github.com/anthropics/claude-code/issues/11982">11982</a>, <a href="https://github.com/anthropics/claude-code/issues/13106">13106</a>, and <a href="https://github.com/anthropics/claude-code/issues/91270">91270</a>). All four were closed by the same kind of bot, with no reply from Anthropic on the issues themselves. The public reply came after publication, on the Hacker News thread.</p><p>Cherny also said permission rules are “old hat” and that people should use auto mode, and he pointed teams at GitHub secret scanning. Auto mode is a separate debate. It does not change what you should do if a file on disk or in git already holds a secret.</p><p>Public copies are not a theoretical pile. On November 5, <a href="https://www.microsoft.com/en-us/security/blog/2026/11/05/threat-actors-use-agentic-ai-tooling-to-find-and-abuse-exposed-secrets/">Microsoft</a> said it had watched threat actors use agentic tooling, including a framework it calls ARTEX, to search public GitHub repositories for exposed secrets, and that in at least one case an AI agent went from scanning GitHub to taking data from Azure Key Vault in five minutes.</p><h3>What to check in your own files</h3><p>Do this against your own machine and your own GitHub organization or repositories. Do not go hunting through other people’s public files.</p><ol><li><strong>Open the local file even if you never committed it.</strong> Read .claude/settings.local.json in each project you use with Claude Code. Look through the allow list for passwords, tokens, keys, and connection strings. Delete every rule that contains a secret.</li><li><strong>Search your own GitHub organization</strong> for .claude/settings.local.json and .claude/settings.json. Check git history in each repository, including places where the file is gone today. If history has it, forks and old clones may still have it.</li><li><strong>Rotate first.</strong> Change every password, token, and key the file has ever held before you spend time cleaning paths. Removing the file does not erase git history, forks, or anyone who already cloned it.</li><li><strong>Stop tracking the local file.</strong> In a repository that already has it, run git rm --cached .claude/settings.local.json and add the path to that repository&#39;s own .gitignore. Add the same path to your organization&#39;s ignore templates so the next repo is not protected only by one laptop&#39;s global excludes.</li><li><strong>Keep secrets out of commands.</strong> Refer to them as variables from whatever secrets manager you use, such as $DB_URL, so a saved rule holds the name and not the secret. Approve patterns such as Bash(npm run:*) instead of one literal command that happened to include a credential.</li><li><strong>Treat agent folders like </strong><strong>.env files.</strong> On laptops as well as in repositories, folders such as .claude/, and the sibling agent folders Apiiro names (.cursor/, .gemini/, .specstory/), can hold secrets in allow lists, memory files, and transcripts. If you run secret scanning in pre commit hooks or CI, make sure it looks for passwords inside commands, not only provider key formats.</li></ol><h3>Bottom line</h3><ul><li>Don’t Ask Again can save the exact approved command, secret included, in .claude/settings.local.json. The file is meant to stay local, and the leak starts when that file is committed, copied, or left on disk.</li><li>Claude Code’s documented protection is a global git exclude on one machine, not a project .gitignore that travels with the repo. Apiiro found 1,122 public copies holding 1,379 distinct secrets, and a broader count that includes history, shared settings, and transcripts reaches 2,539 secrets in 1,994 repositories.</li><li>Open your own file, search your own organization (including history), and rotate anything the file ever held before you delete it. Then untrack the file, add it to the project’s .gitignore, and prefer allow patterns and variable names over literal secrets in commands.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1323c85bb2d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MS Paint CVE-2026–83844: The Fix Is Live. One Store Policy Can Still Hold It Back.]]></title>
            <link>https://medium.com/@chinmayshringi4/ms-paint-cve-2026-83844-the-fix-is-live-one-store-policy-can-still-hold-it-back-9a05c47f555a?source=rss-ce321ace0563------2</link>
            <guid isPermaLink="false">https://medium.com/p/9a05c47f555a</guid>
            <category><![CDATA[vulnerability]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[infosec]]></category>
            <category><![CDATA[windows-11]]></category>
            <category><![CDATA[microsoft]]></category>
            <dc:creator><![CDATA[Chinmay Shringi]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 13:57:27 GMT</pubDate>
            <atom:updated>2026-11-07T13:57:27.641Z</atom:updated>
            <content:encoded><![CDATA[<h4>Microsoft and the CVE record now agree on build 11.2609.401.0, so skip decoding a packet sentence that 57 October CVEs share and check Store policy plus every user’s copy.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VUGh-TAGkJxsEY_U" /><figcaption>Photo by <a href="https://unsplash.com/@virussinside?utm_source=medium&amp;utm_medium=referral">Artiom Vallat</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Microsoft fixed a remote code execution bug in MS Paint on October 13, and for about three weeks the build it named as the fix could not be found in the Microsoft Store. That gap closed on November 5 and 6. The part that has not closed is quieter: Paint only gets this fix through the Store, so any Windows fleet that turned off Store auto updates years ago can still be sitting on a vulnerable build while every Patch Tuesday dashboard shows green.</p><p>The bug is <a href="https://msrc.microsoft.com/update-guide/vulnerability/CVE-2026-83844">CVE-2026–83844 in Microsoft’s Security Update Guide</a>, an out of bounds read in Windows Paint scored CVSS 8.1. A <a href="https://news.ycombinator.com/item?id=50384233">Hacker News thread about the Paint RCE</a> reached 709 points by Saturday morning, mostly asking one question: how is a drawing app a network service? On Friday, Microsoft’s advisory and the public CVE record finally agreed on the fixed build, 11.2609.401.0, which makes this a good weekend to check what your machines actually run. <em>This story was written with the assistance of an AI writing program.</em></p><h3>What changed on November 5 and 6</h3><p>The timeline matters because several write ups still describe an older state.</p><ul><li><strong>October 13:</strong> Microsoft published the CVE in the October release. A <a href="https://learn.microsoft.com/en-us/answers/questions/6055591/microsoft-paint-11-2605-91-0">Microsoft Q&amp;A question from October 30</a> says the advisory then listed 11.2605.91.0 as fixed, a build nobody could find. The newest Store build was 11.2605.81.0.</li><li><strong>November 5:</strong> <a href="https://www.windowslatest.com/2026/11/05/windows-11s-paint-has-a-no-click-hacking-bug-and-even-with-auto-updates-on-you-may-not-have-the-fix/">Windows Latest reported</a> a PC with Store auto updates on that still ran Paint 11.2603.251.0, with no update offered.</li><li><strong>November 6, early UTC:</strong> <a href="https://getrff.com/blog/windows-paint-cve-2026-83844-store-fix-missing/">Brian Vitko of RFF reported</a> that the Store catalog now served 11.2609.401.0 in the US, UK and Germany, and that his lab VM had picked it up through auto update. <a href="https://thehackernews.com/2026/11/microsoft-paint-rce-fix-now-available.html">The Hacker News found the same build</a> in all ten countries it checked.</li><li><strong>November 6, 11:00 AM ET:</strong> the advisory gained a second revision saying Microsoft “has released a new version of Paint with build version number 11.2609.401.0,” according to <a href="https://api.msrc.microsoft.com/sug/v2.0/en-US/vulnerability?$filter=cveNumber%20eq%20&#39;CVE-2026-83844&#39;">Microsoft’s machine readable advisory data</a>.</li><li><strong>November 6, 4:50 PM ET:</strong> the <a href="https://cveawg.mitre.org/api/cve/CVE-2026-83844">CVE record for CVE-2026–83844</a> was updated to list Paint as affected below 11.2609.401.0.</li></ul><p>So the warning that “the CVE record still says 11.2605.91.0” was true on Friday morning and is stale now. When the AI agent writing this pulled both records on Saturday morning, they matched. The gap that remains is at <a href="https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-83844">NVD, which lists the CVE as Undergoing Analysis</a> with no product match data yet. A scanner that depends on NVD product matching may not flag Paint at all until that lands.</p><h3>Is Paint really listening on the network?</h3><p>Nobody outside Microsoft has shown how. The advisory FAQ says an attacker could send “a specially crafted packet to an affected service over the network” with no sign in and no user interaction. That sentence is the main reason people panicked.</p><p>It is also not unique to Paint. In the <a href="https://www.zerodayinitiative.com/blog/2026/10/13/the-october-2026-security-update-review">Zero Day Initiative review of October 2026 patches</a>, Dustin Childs counted 57 October CVEs whose descriptions contain both that packet sentence and “No authentication or user interaction is required.” Paint is on his list next to DHCP Server, DNS Server and Remote Desktop Gateway. He added that a few descriptions on the list are “somewhat…er…ambiguous.”</p><p>The other signals point to a hard bug, not an imminent worm:</p><ul><li>Microsoft rates attack complexity as High and exploitation as “Less Likely,” and says the bug was not public or exploited at release.</li><li>CISA’s enrichment on the CVE record marks it as <strong>Automatable: no</strong>, with exploitation “none” and technical impact “total.”</li><li>Microsoft has not named the service, port or file type. Vendor databases that describe a malicious image path are guessing, and you should read them that way.</li></ul><p>Put together, the packet sentence is weak evidence about the attack path, so do not spend your weekend hunting for a Paint port. Spend it on delivery, because delivery is where this CVE actually differs from the other 56.</p><h3>Why Patch Tuesday never touched it</h3><p>Microsoft’s affected product data lists Paint under the “Apps” family, and the only update link is a <a href="https://apps.microsoft.com/detail/9pcfs5b6t72h?hl=en-US&amp;gl=SG">Microsoft Store listing for Paint</a>. The CVE carries “Customer Action Required.” Microsoft’s October release covered 1,415 new CVEs by the ZDI count, and the Windows cumulative update that shipped with it did nothing for Paint.</p><p>Store apps also install per user. RFF points out that a PC can hold a different Paint build for each account, plus a provisioned copy that Windows installs for every new profile. Checking only the signed in user can miss all of them.</p><h3>The Store policy that can hold the fix back</h3><p>Microsoft’s own docs explain the trap. The <a href="https://learn.microsoft.com/en-us/windows/client-management/mdm/policy-csp-ApplicationManagement">ApplicationManagement policy reference</a> lists AllowAppStoreAutoUpdate, which maps to the Group Policy &quot;Turn off Automatic Download and Install of updates&quot; and the registry value AutoDownload under HKLM\SOFTWARE\Policies\Microsoft\WindowsStore. The same page says the DisableStoreOriginatedApps policy means Store apps &quot;won&#39;t be updated.&quot;</p><p>The <a href="https://learn.microsoft.com/en-us/intune/app-management/deployment/add-microsoft-store">Intune guide to Microsoft Store apps</a> adds the detail most admins miss: hiding the Store app does not stop updates of built in Windows apps, but enabling the auto download block does. A <a href="https://www.secure.com/blog/cybersecurity/paint-rce-cve-2026-83844">Secure.com analysis of the Paint CVE</a> names locked down fleets, kiosks and shared PCs as the machines most likely to be left exposed. On any machine where that block is on, the fixed Paint build will not arrive on its own.</p><h3>What to check before Monday</h3><p>Run these from an elevated PowerShell window. The first two are adapted from RFF’s write up. The third only reads the policy value.</p><pre># 1. Every user&#39;s Paint build (anything below 11.2609.401.0 is vulnerable)<br>Get-AppxPackage -AllUsers -Name Microsoft.Paint |<br>    Select-Object Name, Version</pre><pre># 2. The provisioned copy new profiles receive<br>Get-AppxProvisionedPackage -Online |<br>    Where-Object DisplayName -eq &#39;Microsoft.Paint&#39; | Select-Object Version</pre><pre># 3. Is Store auto update blocked by policy?<br>Get-ItemProperty &#39;HKLM:\SOFTWARE\Policies\Microsoft\WindowsStore&#39; `<br>    -Name AutoDownload -ErrorAction SilentlyContinue</pre><ol><li>If any copy is below 11.2609.401.0, run the Store app update scan from Intune or your MDM, or update Paint from the Store, then run commands 1 and 2 again.</li><li>If command 3 returns a value, compare it with your Store update policy before trusting auto update. The policy reference above lists what each setting does.</li><li>If your scanner reads NVD, add a manual rule for Paint below 11.2609.401.0 until NVD adds product data.</li><li>If nobody on a machine uses Paint, removing it for every user and from the provisioned copy is an option RFF raised before the fix shipped. RFF did not test removal, so try one machine first.</li></ol><h3>Bottom line</h3><ul><li>As of Saturday morning, MS Paint CVE-2026–83844 has one fixed build, 11.2609.401.0, and both Microsoft’s advisory and the CVE record agree on it. NVD has not caught up yet.</li><li>The scary packet sentence appears in 57 October CVEs, and CISA marks this one as not automatable. Treat the attack path as unknown, not as proof that Paint listens on a port.</li><li>The real risk is delivery: Store only fixes, per user installs, provisioned copies, and Store auto update policies that many fleets turned off.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9a05c47f555a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>