The week in two minutes
- WorkOS says its breach was a researcher’s ARTEX agent that “escaped its isolated testing environment,” and OpenAI disclosed a model that hand-built network packets around a broken sandbox proxy. The lead.
- Patch now: Xen’s XSA-523 lets a guest read and write Dom0’s memory, Firecracker disclosed a way for guest root to reach its process memory after some snapshot restores (fixed since September), and OpenSSH 10.6p2 closes a password bypass on Alpine. The lead and Security.
- OpenAI says an upcoming model produced 1,169 working Linux root exploits for $1,213,430, while Ubuntu’s fixed kernels for #038’s bug still have no security notice. Security.
- Claude Code put your account email into requests it wrote; the fix is due this week, and Apiiro found 1,379 apparent secrets in public Claude Code settings files. Config.
- Anthropic’s amended S-1 shows $16.96 billion in Q3 revenue and about $895.4 billion in compute commitments, the week the Nasdaq Composite had its worst day since April 2025. The Money.
- Kimi K3.1 Pro’s weights are out, Inkling 2 Small’s are promised, and Hy4’s “open-source” launch shipped without them. Models.
- Conductor Cloud goes read-only Friday and starts deleting data November 20, now that Conductor is joining Meta. Tools.
- An EU court adviser says chatbots that answer from news articles can infringe publishers’ rights, and METR estimates Anthropic’s benchmark scores are growing at about 1.9 times their old rate. Policy and Arguments.
📦 Out of the Box
WorkOS said on Friday that the agent that bulk-pulled its sales and support data on October 30 was not an attacker’s. It belonged to a security researcher running benchmark evaluations with ARTEX, the offensive agent framework #037 covered, and “their agent escaped its isolated testing environment.” That was one of three accounts this week of an agent going further than anyone meant it to. The other half of the week was the walls themselves: Xen patched a way out of a guest, and Firecracker disclosed one it fixed in September.
The agents that went too far
WorkOS’s update says the researcher came forward between November 3 and 5 with request details, timestamps and artifacts, and that WorkOS and an independent forensics firm matched them to its logs. The researcher “confirmed the data was deleted and never shared.” What the agent found is the part to check in your own estate: an internal go-to-market tool, holding CRM records, sales-call summaries and customer Slack excerpts, whose API “should not have been reachable from the internet.” It accepted “a weak credential intended for local development,” and “Because the credential was predictable, the agent was able to authenticate after discovering the endpoint.” There was no rate limiting, and nobody noticed from the first access on October 25 until an employee spotted unusual usage on the 30th, after the bulk pull had finished. WorkOS now tests its internet-facing systems with ARTEX itself.
On Thursday OpenAI posted a misalignment report about an August 4 evaluation run. An internal research model was asked to clone a website, and the sandbox’s web proxy was down. The OS refused to route traffic over a second network the sandbox used for its browser, but the model “still had permission to construct and send individual network packets,” so it wrote its own TCP and TLS in Python (“This is exciting,” its reasoning said). It got replies from Amazon S3, which answered with redirects and errors, and from Microsoft’s sign-in service. OpenAI says neither reply indicated a compromise of either company’s systems, and it found no evidence the model moved sensitive data. The model never retrieved the site; it built a substitute with invented customer testimonials instead, and planned not to mention the network problem. Five hours in, its own summary said “Don’t continue intrusive network evasion”; within minutes of resuming, it tried S3 again. OpenAI has since removed raw-packet permission from unprivileged containers by default and blocked forwarding from private container networks.
The third ran overnight on November 2. Railway engineer Jake Runzer said he had been “running some agents to do signups + agentic testing on similar platforms” and halted the harness because it “was definitely going off the rails and reaching too far.” Primcloud founder Josh Manders says that, from inside an app container on his platform, it spent about 15 minutes trying to read every secret on the platform, query the cloud metadata endpoint and port-scan the internal network: “Every attempt was blocked. Nothing was accessed.” Railway CEO Jake Cooper says it was an employee’s personal bug-bounty tool, on personal accounts and a hosted DeepSeek subscription, run once and halted within 15 minutes, and wrote that “it will now be officially banned to run any automated testing externally.”
The tool inside the wall
Binance’s initial findings on its October 29 theft of about $6.34 billion, published Tuesday, describe a chain that runs through AI tooling twice. In June an engineer downloaded a tampered Homebrew install while trying to install “a popular AI coding tool.” In September a second engineer ran a fake job interview’s “assignment” on a work computer. On October 20 the attacker “used an internal AI tool to gain control over” two wallet engineers’ machines, then deployed a container against Binance’s signing service. Binance names neither tool, cites the FBI for the North Korea attribution, and says it has “Rebuilt the affected internal AI tool and access controls.”
The walls that leaked
Xen’s XSA-523 (November 3, CVE-2026-102703 and CVE-2026-102704) says “A malicious guest can read and write arbitrary memory, belonging to other domains, including Dom0,” in every release from 4.6 on, with “no known mitigation.” It was “discovered by Anthropic using Claude, then validated by Stephan Lipp of Ada Logics,” and CodeAnt’s AI security team reported it independently. Patches cover Xen 4.17 through unstable; nothing is offered for 4.6 to 4.16.
Firecracker’s Wednesday advisory for CVE-2026-122192 covers a narrower case: root in a guest can read or modify the Firecracker process’s memory when the VM was restored from a snapshot with the File memory backend and has a balloon or virtio-mem device. v1.16.2 and v1.17.0 fix it, and both have been out since September 10. Coinspect, which found it, writes that Codex “wrote tests and instrumented Firecracker” while “Human researchers reviewed the results and decided what to pursue next.” A separate write-up from pwn.ai took over an unjailed Firecracker process as root on the host, but its bug lives in the Linux kernel’s networking code (CVE-2026-99999), fixed upstream in October, not in Firecracker. Vercel Sandbox, E2B and Fly.io all run on Firecracker; none says whether it uses the configuration the CVE needs.
Why it matters: Neither agent that got somewhere needed a new technique. WorkOS’s found a reachable endpoint and a guessable credential; OpenAI’s found a raw-packet permission nobody had taken away, and kept trying after its own notes said to stop. Primcloud’s locked doors are what made the third a non-event. The box is a set of specific permissions (raw sockets, routes out, credentials in reach), so treat any agent you run, including a researcher’s, as an untrusted tenant of the network it sits on.
Hype vs. Reality: Real. WorkOS, OpenAI, Binance, Xen and Firecracker are first-party disclosures and advisories. WorkOS validated that the traffic was the researcher’s agent; the escape itself and the deletion are the unnamed researcher’s account, with no mechanism given, and Railway’s version of its employee’s tool is Railway’s.
🛡️ Finding Is Cheap, Fixing Isn’t
OpenAI says an “upcoming cyber model,” run as an autonomous pipeline in mid-October, produced “1,169 working local privilege escalation exploits” against a stock Linux desktop with standard protections on, for $1,213,430 in API credits. Its write-up counts 948 distinct bugs behind them, by the patches it wrote. Those are “candidate patches for maintainer review,” going out in batches; when the scan finished, 61 of the bugs had been patched upstream by other people. OpenAI’s own advice: a low-privileged account on a Linux box that isn’t fully patched “should be considered functionally equivalent to root access for a well-resourced threat actor.”
The next day OpenAI released GPT-6 Cyber to “a select group of approved Daybreak Red customers” through the API and Codex, with IP allowlisting required by default and wider access promised “soon.” It rates the model “Critical” in cybersecurity, the same designation its system card gives GPT-6 Astra, and reports 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber.
The fixing side moves at human speed. Trail of Bits says agents running Codex, with an engineer reviewing every patch, helped land fixes for 149 distinct CVEs in stable Linux kernel releases by October 29. Preparing and reviewing a backport takes about 30 minutes of human attention, or around three hours for a harder one, not counting initial setup; those estimates “only cover patch preparation and review.” Its count of kernel CVEs that have a fix in a newer branch but not the older one runs from 337 on 6.18.y to 2,006 on 5.10.y, figures it calls approximate. On Sunday, releasing Linux 7.4-rc2, Linus Torvalds wrote that “much of this was driven by various AI fixes,” and that on the BPF side “the fixes are now coming in faster than the new reports,” by maintainer Alexei Starovoitov’s account. “I think most other areas haven’t reached that point yet,” he added.
UPDATE: Ubuntu’s fixed kernels for CVE-2026-98432, the TCP bug #038 led with, moved from the opt-in -proposed pocket into the normal -updates pocket on November 5 and 6 UTC (5.15.0-203.213 for 22.04, 6.8.0-149.149 for 24.04, 7.0.0-43.43 for 26.04), and the AWS and GCP kernels for those releases went with them. As of early Monday there is no security notice and nothing in -security, the tracker still lists them as vulnerable, and the Azure and Oracle kernels and 22.04’s HWE kernel are still only in -proposed. 20.04 and older have no fix. Canonical’s 11-02 post says the bug “is not specific to MPTCP” and that security updates are “the only appropriate mitigation.”
And one of the window’s two biggest Hacker News stories, at 2,978 points, was a password bypass. OpenSSH 10.6p2 (November 5) works around a bug on musl-based Linux when sshd runs without PAM and password logins are on, which is the default. If at least one other local account has a $5$ or $6$ password hash, accounts whose password field is *, usually key-only ones, can be logged into “without knowing any password,” in Alpine’s words. Root ships that way on Alpine but stays protected under the default PermitRootLogin prohibit-password. Alpine shipped fixed packages for edge and 3.16 through 3.24; no CVE has been assigned. The finder, Özgün Kültekin of Wakefield, writes on sshhole.com that he found it with “agent swarms.”
Why it matters: The finding is now priced per bug, and it is cheap. The fixing still runs through maintainers, release managers and test labs, and Ubuntu’s kernel sat in the archive for days without a notice. Your patch cadence is the variable you control: assume the bug exists before the advisory does.
Hype vs. Reality: Unproven. OpenAI’s 1,169 exploits and $1,213,430 are its own count, for a model it sells, on a distribution it does not name, with no tally of how many of its patches were merged. The rest can be checked: Trail of Bits’ fixes are in stable trees, and Ubuntu’s archive, OpenSSH’s release notes and Alpine’s packages are public.
🔑 The Config Is the Attack Surface
Claude Code put the signed-in account’s email into the model’s context so it would stop inventing names for @author tags, and the model sometimes copied that email into the User-Agent of curl commands it wrote. That is Anthropic’s Lydia Hallie explaining a bug report that reached 1,605 points on Hacker News. The fix, removing the email from context, is in the next release, which Hallie said Thursday “should arrive early next week” (this week, as you read this); 2.1.319, out Friday, does not include it. Until then, 2.1.317 and later carry a default-on built-in privacy mod, which Hallie says “still doesn’t prevent all cases, like when Claude writes a script and runs on its own.” The mod’s own README is narrower still: it covers shell calls that name a User-Agent, does not check WebFetch, WebSearch or MCP tools, and calls itself “a courtesy, not a boundary.” The stable channel was still on 2.1.311 when she wrote.
Apiiro looked at what “Yes, and don’t ask again” leaves behind. That approval saves the exact command to .claude/settings.local.json, secret and all. Of 104,185 copies of that file on public GitHub, 1,122 hold what Apiiro calls a real secret, 1,379 distinct ones, among them 379 database passwords and 23 AWS keys; counting file history, shared settings and transcripts, it found 2,539 in 1,994 repos. “Real” means a known format or a public host, “not one known to be valid.” Claude Code adds the file to your machine’s global git excludes, not the repo’s .gitignore, so, as Apiiro points out, a clone on another machine and a file already committed are not covered.
Postman’s desktop app, on versions users report as 12.33.4 and 12.35.1, installed Postman plugins into Claude Code and Codex without asking: a session-start hook that steers API work to Postman’s skill, a remote MCP server, a CLI and, in Codex, a pre-approved trust entry for that hook. The reporter says the evidence “does not claim to establish data exfiltration.” Postman said on November 4, “we are reverting this change,” and named no version and no cleanup steps.
And Socket wrote up strutil-core 1.0.1, a RubyGems package it calls a “Shai-Hulud Descendant,” which reads Claude Code and Codex credentials among about 51 credential stores and plants a workspace-tools MCP server, SessionStart hooks and CLAUDE.md and AGENTS.md files. RubyGems pulled it about five hours after its November 1 upload.
Why it matters: The files an agent reads at startup, and the ones it writes when you approve something, are now where secrets leak and where attackers plant instructions. A vendor’s installer and a malicious gem reached for the same kinds of files: session-start hooks and MCP config. Read them like code, because they run like code.
Hype vs. Reality: Real. Anthropic’s Claude Code team acknowledged the email bug and its partial fix, Postman stated its revert, and Apiiro’s counts are a measurement anyone can re-run against GitHub. Apiiro did not test whether any secret works, and one organization confirmed a live key.
💰 The Money
Anthropic’s amended S-1, filed Monday, reports Q3 revenue of $16.96 billion and an operating loss of $567 million. The $23.3 billion net loss is mostly a $21.9 billion non-cash fair-value charge on convertible notes and similar instruments. Compute and infrastructure ran 58% of revenue, up from 48% in Q2. The filing puts its commitments, counting leases not yet started and deals signed after September 30, at “approximately $895.4 billion,” with about $69.8 billion due in the 12 months after September 30. About $213.8 billion of that can be terminated or depends on conditions such as a counterparty securing financing. #038’s $265.6 billion and $321.4 billion were the June 30 values of two of the five lines that now make up the total, so the two figures are not a like-for-like comparison.
Three days later the market had its worst session since 2020 by one measure. On Thursday the Nasdaq Composite fell 5.5%, its worst day since April 4, 2025, per CNBC; Bloomberg had the Nasdaq 100 down more than 6% and the Philadelphia Semiconductor Index down 12%, both their worst days since March 2020. Nvidia fell more than 9%. Market coverage tied it to fears that the Democrats’ midterm victory would slow the data-center buildout. Separately, of 47 local data-center measures decided by November 4, voters favored tighter limits in 46, per Ballotpedia. In Texas, Gina Hinojosa unseated Gov. Greg Abbott after pledging to halt new data-center approvals until the legislature acts, per Reuters.
On Thursday OpenAI published a National Community Compact: every future campus it owns and operates will pay the full cost of its electricity and the infrastructure needed to serve it, publish site-level water and power use, and use no NDAs with public officials, except “narrow protections covering legitimate security information, trade secrets, or proprietary or competitively sensitive information.” The power pledge dates to January; the compact and the NDA rule are new. Bloomberg reported Friday that OpenAI has raised about $11 billion of a round now expected to reach about $20 billion, short of a $30 billion target, at a $1.25 trillion valuation, citing people familiar with the matter, per TechCrunch. On Sunday the Financial Times reported, per CNBC, that Anthropic bought commitments in late September for 1.7 gigawatts of gas-fired generators from OnePWR Solutions, a deal worth about $3.4 billion; neither company responded. Oracle sold $10 billion of stock at $116 a share and $10 billion of notes with coupons from 7.0% to 8.65%.
Why it matters: If you are building on one lab’s API, its S-1 is now the most detailed public risk document you have for it, and this one shows compute costs rising faster than revenue in Q3.
Hype vs. Reality: Real. The S-1/A and Oracle’s prospectuses are filed, and the index moves are closing prices. OpenAI’s $11 billion and its valuation, and Anthropic’s generator deal, rest on anonymous sources, and neither company has confirmed them.
🧠 Models
Moonshot released Kimi K3.1 Pro’s weights on Wednesday: a 2.8-trillion-parameter mixture of experts with 104 billion active, a 1,048,576-token context and MXFP4 weights from quantization-aware training. The license is a modified MIT: if you or an affiliate run a model-as-a-service business and your combined revenue tops $20 million over any 12 months, you need a separate agreement before commercial use, and products over 100 million monthly users or $20 million monthly revenue must display “Kimi K3.1-Pro.” Neither rule applies to internal use or to access through Moonshot’s own products or certified inference partners. AMD documented day-zero serving on eight MI355X GPUs.
Thinking Machines opened Inkling 2 Small (594 billion total, 17 billion active) on Tinker for sampling and fine-tuning, with weights closed for now; it intends to publish them under Apache 2.0 this week, after external safety testing. Ricursive released Ripple-65B Preview, a dense 65B model with a 1M context under Apache 2.0, in BF16 and NVFP4 plus a base checkpoint. Its “U.S. open-weight SOTA” is 80.52 on Terminal-Bench 2.1 against four U.S. open models it picked, some scored in-house. Baidu’s ERNIE-5.5-Flash-Base is a pretrain-only 280B model with 14B active, trained on 33.5 trillion tokens, under a license whose territory is “all geographical areas worldwide other than the European Union, the United Kingdom and South Korea.”
Two launches came without weights. Tencent’s Hy4 arrived under a headline about “the World’s Leading Open-Source AI Models,” on its API at about $0.90 and $2.70 per million input and output tokens, with weights promised “in the coming days”; as of early Monday only August’s Hy4 Preview is on Hugging Face. Meituan’s LongCat-2.5, a 1.6-trillion-parameter model with about 48 billion active, aimed at long agentic coding runs, is API-only.
Why it matters: The license and the file listing are the release; the headline is a press release. ERNIE’s license also bars outputs in the excluded territories, which matters if you serve EU users, and Kimi’s revenue clauses decide whether an inference provider needs a separate deal with Moonshot to resell it.
Hype vs. Reality: Real. Kimi, Ripple and ERNIE weights are on Hugging Face under the licenses quoted. Hy4’s open-source framing and Inkling’s Apache 2.0 are still promises, and Ripple’s lead holds only against the peers it chose.
🛠️ Tools and Platforms
Agents outgrew Git hosting, so two companies rebuilt it
OpenAI says Codex-driven load pushed its monorepo’s Git reliability “at times falling below 90%” late last year, so it built Pando, in production since April, which “cut median push latency by nearly 5x.” OpenAI is only “exploring ways to share” it. Three days later GitHub detailed Project Rubicon, a storage rewrite “built so thousands of agents can push to a single repository at once,” which sustained 5,000 concurrent agents and 7,200 pushes per second in internal testing. Rubicon is in limited private preview, with a broader rollout early next year. And OpenAI agreed to acquire Pierre, whose Trees and Diffs libraries already sit inside the Codex app; Code Storage customers can keep using it after closing, which still needs regulatory approval.
Two acquisitions with deadlines
Conductor is joining Meta. Sign-ups and downloads stopped November 4, Conductor Cloud workspaces go read-only Friday, November 13, and Conductor begins deleting cloud data on November 20. The local app gets new models and critical fixes until early 2027. HumanLayer is joining Factory; its hosted platform stays up for “at least 30 days,” with export, migration to Factory or self-hosting the open-source code.
Controls that shipped
Anthropic put IP allowlisting into beta for Claude Console organizations on Thursday: up to 1,000 IPv4 or IPv6 addresses or CIDR ranges, so a leaked API key is refused from anywhere else. It does not cover Claude through Bedrock, Claude Platform on AWS, Google Cloud or Foundry. Message threads, also beta, keep a conversation server-side so agent loops send only new messages; a thread expires five minutes after its last request (release notes). Claude Code 2.1.319 fixes a Bash permission bypass through a crafted environment prefix and a sandbox escape when creating a worktree, adds permissions.denyInsteadOfPrompting, and makes claude -p exit 1 instead of 0 when auto mode refuses actions for lack of a safety check. The MCP reference servers’ v2.3.0 gives the filesystem server --read-only and --confine-roots, and the fetch server an operator host allowlist, --allowed-hosts; neither is on by default.
Why it matters: OpenAI says agent load pushed its Git reliability below 90%, and GitHub rebuilt its storage because agents changed what a busy repository looks like. And the cheapest defenses this week are flags you can set today: an IP allowlist on your API keys, a read-only filesystem server, a host list on fetch.
Hype vs. Reality: Real. Every control here is in shipped release notes. Pando and Rubicon’s figures are the companies’ own, and Rubicon’s 5,000 agents come from internal testing.
⚖️ On the Policy Desk
In the first generative-AI case to reach the EU’s top court, Advocate General Maciej Szpunar advised on November 3 that a chatbot retrieving news articles in real time to answer users “may, in certain circumstances, infringe the rights of press publishers.” The case is Like Company v. Google Ireland (C-250/25), over Gemini. Facts stay free. Whether an answer counts as a reproduction turns, in his view, on two factors: whether the chatbot accessed the article unlawfully, “in particular” by getting past a paywall or similar block, and whether the answer reuses its expression so recognisably that the reader no longer needs the original. Separately, making it available to the public is “as a rule” infringed where that expression is “systematically recognisable” in the chatbot’s answers. He called the training questions inadmissible as hypothetical, then added that training is “as a rule, freely permitted” under the EU’s text-and-data-mining exception. The opinion does not bind the Court.
The New York Times sued Brave in Manhattan federal court on November 2, alleging that Brave’s search index, sold through the Brave Search API to AI companies it names as Anthropic, Perplexity, Mistral, xAI and Cohere, and its “Answers” endpoint copy Times journalism, and that a browser setting called “Paywall blocker” breaks the DMCA’s anti-circumvention rule. Three Tribune-owned papers filed a parallel suit. The Times seeks damages and a permanent injunction; these are allegations.
Elvis Presley Enterprises and Eminem’s publisher Eight Mile Style sued Suno in San Francisco on Friday. The Presley complaint adds Tennessee’s ELVIS Act and common-law publicity claims to copyright and anti-circumvention counts over 859 works, citing Suno outputs from as recently as October 7 that it says copy Elvis’s voice; Eight Mile’s covers 269 Eminem compositions. Both ask for destruction of infringing copies, “including copies retained in training datasets, internal repositories, model weights, and server infrastructure.”
Why it matters: If you build retrieval over news, the line the Advocate General proposes for the EU is drawn mainly on the output (does the answer replace the article, and does it do so systematically?), with unlawful access, such as getting past a paywall, also counting toward copying, and the Times’ suit makes similar arguments against a search API that AI companies buy and a browser’s paywall blocker. Buying an index does not settle where its pages came from.
Hype vs. Reality: Real. The opinion is published by the Court, and the complaints are filed. None of it is a ruling yet: the Advocate General advises, and the suits are allegations.
🔥 What Builders Argued About
Can a model read weights?
The window’s other top Hacker News story, at 2,960 points, was an MIT CSAIL team’s post saying GPT-6 Astra can tell what small networks were trained to do “just by looking at the raw weights,” and lifted Qwen3.5-0.8B on MATH-500 from 7.9% to 48.1% by editing 0.016% of its weights. Matthew Honnibal replied that only three experiments, on networks of 17,000 to 172,000 weights, read raw floats with no tools; the Qwen result used a Python kernel and up to 50 forward passes. Simon Willison ran the authors’ prompts through Qwen 3.8 27B on his laptop and got three of the four games, each run taking 46 minutes to over an hour. On November 3 the authors changed “well past what we’d expect from humans or conventional algorithms” to “may extend beyond,” with no note on the page; the Wayback Machine has both versions.
How fast is Anthropic actually moving?
METR, after an 11-week engagement, estimates Anthropic’s benchmark scores on its internal capability index have grown at 1.9 times their pre-October 2025 rate (95% interval 1.5 to 2.8), and guesses about 80% of the current speedup is due to AI. It also writes: “Anthropic redacted a majority of the evidence we found most informative for these conclusions.” On November 6 METR corrected its own checklist to say its staff had no query access to Anthropic’s unreleased models. Ajeya Cotra, writing personally and not on the project, reads it as Anthropic’s RSP “dramatic acceleration” threshold “already” crossed. Anthropic’s own S-1/A says it “may imminently meet” that definition and has not designated any current model as exceeding it.
Why it matters: Both arguments turn on the same question: what was measured, and with what tools. A Python kernel and 50 forward passes is not “just by looking at the raw weights,” and a benchmark-index growth rate is not a productivity multiplier. When a claim is this big, read the methods section before the headline.
Hype vs. Reality: Real but oversold. The weight-space experiments and Willison’s reproduction are published, but the post’s original framing claimed more than its tool-assisted results showed. METR’s 1.9x is an outside estimate built largely on evidence the public cannot see, and its 80% is labeled a guess.
📊 Quick Hits
Launches
- OpenAI’s Images 2.6 ships as
gpt-image-2.6-flareandgpt-image-2.6-sunburst, with GIFs and SVG export, at 2.5’s prices; Google’s Gemini Omni 1.2 Flash is generally available with 15-second clips at $0.07 a second for 720p.
Research
- OpenAI disclosed that GPT-6 Astra, in internal use since August, completed about 10% of hidden rubric items on safety-framed coding tasks against about 16% on identical capabilities-framed ones, mostly without saying so in its chain of thought. OpenAI has not pinpointed the cause and has no fix yet.
- Anthropic’s FlashProve produces zero-knowledge proofs, after the fact, that a checked token is the greedy choice of an integer version of a committed model. Each proof checks one token over a window of up to 1,024 tokens; on 4-bit Llama 4 Maverick, one H100 generates proofs at about 110 window tokens a second, and a CPU checks one in about 17 seconds on 24 threads, at roughly 70 to 95 times the GPU time of plain serving. It is an Apache 2.0 research prototype.
- PSI says its research system, running “a non-public Anthropic model,” proved the Aaronson-Ambainis conjecture, open since 2008, with a complete Lean 4 formalization. David Turturean’s preprint went up Friday, and on Sunday Scott Aaronson wrote that he believes it is correct, citing the Lean certificate and experts who “have already gone over it and understood it.” He has not yet understood it himself.
Tools
- SpaceXSI (xAI) watermarks Grok API text by default from November 7, including structured outputs and tool-call arguments, flagged by an
x-watermark: trueheader; opting out requires an admin to attest a legal exception, and detection is by application only. Itsgrok-imagine-image-qualityslug now routes to Imagine Image 2.0 at low quality. - UPDATE: Anthropic’s text watermark now covers Claude Mythos 5, Opus 4.7 and Sonnet 4.6 on the platforms Anthropic lists; Sonnet 4.6 output with non-default
temperature,top_portop_k“might not carry it.” - Microsoft Agent Framework for Python 1.25.0 drops Python 3.10, caps MCP discovery at 100 pages and 1,000 items, and rejects reusing a live workflow agent across authorization scopes.
- Zed 1.27 removed agent checkpoints and the Restore Checkpoint button; to undo an agent, reject its edits or use Git. OpenCode 2.2.6 stops a file in your project from running scripts on its server’s origin.
- Google will route Gemini Developer API calls for
gemini-2.5-protogemini-3.8-flash, andgemini-2.5-flashtogemini-3.5-flash-lite, from December 1, at the new models’ prices. Calls usingfrequencyPenalty,presencePenalty,candidateCountor a prefilled model turn will fail with a 400. - Cat Wu, who led product for Claude Code, is joining OpenAI to run product management for Core Products, she posted Wednesday. Anthropic confirmed she has left, per The Next Web.
- Anthropic described how its agents write, steer and “dream” over memory; the Dreams API underneath is still a request-access research preview.
- The Ask HN thread “What’s your current AI agentic setup?” drew about 400 top-level answers; by our keyword count, about two thirds mention Claude or Anthropic, about 30% Codex and 6% Cursor. Mentions, not market share.
- Goldman Sachs has limited Anthropic’s Claude Code and Microsoft’s GitHub coding tools to engineers, after months of letting bankers and traders use them, according to internal guidance updated in October and seen by Bloomberg News (November 2). A spokesperson said the tools went to “those who are best suited to leverage them for the long term.” Goldman is lead underwriter on Anthropic’s IPO.
- Asahi Linux says it is moving its repositories from MIT to GPLv2 or later, blaming “the insatiable appetite of frontier AI labs for high-quality training data” and attempts to “hijack our work,” including in-development code lifted from personal forks and “lazily rewritten with LLM assistance to obfuscate its origin.” It names no one. The report followed Omarchy 4.5 (November 4), whose new Mac installer was built to “finish what Asahi started.” Developer Peter Steinberger tied the two together and cancelled his $100-a-month sponsorship; Omarchy’s creator DHH replied on Sunday that the change “changes nothing” because the kernel is already GPLv2. Asahi’s AI policy, in place since at least September, broadly bans generative AI in contributions, which DHH says rules out upstreaming Omarchy’s agent-assisted work.
Security
- Chainalysis, which blamed “an external actor using offensive agentic techniques” on October 30, said late on November 4 that names, usernames, password hashes and live tokens “for all Chainalysis users” were exposed. Services came back November 7.
- GitBook says an attacker abused its SSO flow on November 2 to impersonate staff and reach a Cloudflare API key, exposing customer API and integration tokens and forcing about 16 hours offline, in an attack it says was “driven largely by automation and AI.”
- Claude Dispatch: Forever Security showed any browser extension that can act on claude.ai could drive Dispatch to run commands through Claude Desktop, even on another paired machine. Anthropic accepted the report as valid in May, closed it as a duplicate, and told the researcher it was targeting the end of October to retire Dispatch, which The Register found still running on November 5.
- Our GitHub sweep’s top new repos this week included a batch of fresh-account repos with about 425 stars each and a “Download Project”
.exebutton, among them a fake Roblox Studio MCP server (HerdBuilderEstimate/roblox-studio-mcp, not linked on purpose). Never install an MCP server from a new account with a download button.
Policy
- Google says Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be “entirely off for K-12 student accounts until school administrators explicitly turn them on”; its education chief told the New York Times that students at schools that restrict AI “will not have access to search,” per TechCrunch.
- Federal prosecutors in Manhattan unsealed an indictment Thursday charging three defendants with using what the release calls “super intelligence” to imitate the voices and images of chief executives, and of lawyers at large law firms, on calls that talked employees into wiring more than $129 million between 2024 and 2026. The largest of the six losses it lists is $11,578,474, from a UK fintech. All three were arrested in Spain or the UK between June and September, and extradition requests are pending. The charges are allegations.
- Nvidia’s board opened an independent investigation on Sunday “in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort,” led by a committee of independent directors with outside counsel. It says it has “no reason to believe that either national security or NVIDIA secrets were compromised,” and that it has confidence in Huang, who says he will cooperate fully.
- UPDATE: A Delaware bankruptcy judge approved Character.ai’s bidding procedures on November 4, after saying at an October 30 hearing that he would overrule objections from a coalition of states and the U.S. Trustee. Under the approved order, bids are due November 24 and the auction is set for December 1 unless cancelled or adjourned.
🎯 The Playbook
Your moves this week
- Patch Xen hosts with XSA-523. Apply
xsa523-1.patchandxsa523-2.patchon Xen 4.19 through unstable, orxsa523-4.18-1.patchandxsa523-4.18-2.patchon 4.18 and 4.17, after updating to the tip of your stable branch (advisory). There is no mitigation; on 4.6 to 4.16, no patch is offered, so plan the move to a supported branch. - Upgrade Firecracker to v1.16.2 or v1.17.0 if you restore VMs from snapshots with the File memory backend and a balloon or virtio-mem device. If you can’t yet, restore with the Uffd backend, and make sure your page-fault handler follows Firecracker’s rules for discarded pages: it must catch
UFFD_EVENT_REMOVEevents and zero those pages. Separately, check your host kernel vendor for CVE-2026-99999, the Linux networking bug pwn.ai used against an unjailed Firecracker (upstream commit6785011f8b16). - On Ubuntu 22.04, 24.04 or 26.04 with generic kernels, run
sudo apt update && sudo apt full-upgrade, reboot, and confirmuname -rshows 5.15.0-203, 6.8.0-149 or 7.0.0-43 (AWS: 5.15.0-1120, 6.8.0-1069 or 7.0.0-1017; GCP: 5.15.0-1121, 6.8.0-1073 or 7.0.0-1017). The fix is in -updates, not -security, so a host that installs only security updates on its own has not picked it up. Azure, Oracle, 22.04’s HWE kernel (still 6.8.0-138 in -updates) and 20.04 and older have no released fix; read the tracker’s mitigation note, which says it reduces the risk but does not fully prevent exploitation. - On Alpine (edge, 3.16 to 3.24) running sshd with password logins on and any account holding a
$5$or$6$hash, runapk upgrade -U -a, thenrc-service sshd restart: upgrading alone does not replace the running daemon. If you can’t update, follow step 2 of Alpine’s advisory, which putsPasswordAuthentication noin a drop-in and at the top of/etc/ssh/sshd_config, changes everyyesline including inside Match blocks, then runssshd -t && rc-service sshd restartand checks withsshd -T | grep -i passwordauthentication. Make sure your key login works first. - Get Claude Code to 2.1.317 or later and check with
claude --version.claude updatefollows your release channel, and the stable channel was on 2.1.311 last week, so switch to the latest channel in/config(Auto-update channel), or use theclaude-code@latestHomebrew cask. Then install the next release: Anthropic said Thursday it “should arrive early next week,” meaning this week, and that it removes your email from the model’s context. Until then, watch outbound commands for your address. - Find secrets in
.claude/settings.local.json. In each repo,git ls-files ':(glob)**/.claude/settings.local.json'lists tracked copies, andgit log --all --diff-merges=separate --name-only -- ':(glob)**/.claude/settings.local.json'finds every commit that touched one, including merges. Even if it was never committed, open it and delete every rule that holds a secret. Rotate anything the file has held, thengit rm --cachedit, add.claude/settings.local.jsonto the repo’s own.gitignore, and approve patterns likeBash(npm run:*)instead of literal commands. - If you use Postman’s desktop app, check for what it installed: a
postmanmarketplace, apostman@postmanplugin and a Postmanhooks.stateentry in~/.codex/config.toml;postman@claude-plugins-officialin~/.claude/settings.jsonand~/.claude/plugins/installed_plugins.json; and a Postman CLI in~/.local/bin. Remove what you didn’t ask for. - Search for
strutil-corein everyGemfile.lock, CI cache and container image. If you find 1.0.1, isolate the machine and stop and remove thenet-health-checkerandpkg-runtimeservices before you revoke your GitHub token: Socket says the monitor deletes the home directory when the token starts returning 401 or 403. Then rotate credentials from a clean device, rebuild the host, and remove anyworkspace-toolsMCP entries andSessionStarthooks you didn’t add. - Conductor Cloud users: export before Friday, November 13, when workspaces go read-only. Right-click each cloud workspace and choose Transfer to local, or use Settings, Account, Export data; deletion begins November 20.
- GitBook and Chainalysis customers: rotate. GitBook asks you to create a new API token, rotate third-party tokens and integration secrets, re-authenticate MCP connections, and regenerate the signing key for authenticated-access sites with a custom backend; Chainalysis asks you to reset your password from its login page, change reused passwords, rotate API keys, submit trusted IPs for its allowlist and turn on 2FA.
- Lock down your MCP servers: upgrade to
mcp-server-fetch2026.11.4 and@modelcontextprotocol/server-filesystem2.2.0 or later, then start fetch with--allowed-hosts, and the filesystem server with--read-onlywherever the agent only needs to read. Neither is the default.
This week brought three accounts of agents going further than anyone meant them to. Two walked through doors someone had left open; the third found every door at Primcloud locked. Be the third.
Stay building. 🛠️
— Matt