The week in two minutes

  • Two labs claimed the Exponential Time Hypothesis is false, hours apart, with Lean proofs. The researchers examining OpenAI’s claim say their involvement “should not be taken as certification of the claim.” The lead.
  • Patch the Linux TCP bug Anthropic’s model exploited. Red Hat shipped fixed CVE-2026-98432 kernels Sunday and Debian in October; Ubuntu still had none early Monday. The lead.
  • Cursor’s marketplace served a backdoored Claude Code build for about 11 hours on Thursday. Security.
  • Anthropic’s S-1 went public: it has applied to list as ANT, with about $265.6 billion of minimum purchase commitments, mostly cloud hosting and infrastructure, plus about $321.4 billion in leases not yet started, and plans one founders’ share with 50.1% of the vote. The Money.
  • The model on top of Artificial Analysis’s index is one you can’t call: Gemini 4 Carbon goes to cyber defenders first, GPT-6.1 Terra costs $1/$5, and Mistral Large 4’s weights are out. Models.
  • More than 40% of GitHub pull requests now start with an agent, and Jules goes read-only after November 11. Tools.
  • Usage limits were the product news: OpenAI’s credits and December token pool, no 5-hour limit for verified Claude Ultra accounts, and a class action over Perplexity’s limits. The meter.
  • New York signed a law barring 12 chatbot features for any user until an age check shows they are not a minor, and its definition reaches general-purpose assistants. Policy.

📡 Proof of Work

On Monday afternoon OpenAI posted a paper claiming the Exponential Time Hypothesis is false, with a Lean proof attached. About six hours later Anthropic posted its own. The hypothesis, ETH for short, is the standard assumption that 3-SAT needs exponential time, and hundreds of papers prove hardness results that assume it. Both labs say an unreleased internal model found the core result. The next day Anthropic’s red team showed what an unreleased model did to Chrome and the Linux kernel, and that half of the week comes with a patch you need.

Two labs, one hypothesis

OpenAI’s paper claims a deterministic algorithm that counts the solutions of any k-SAT formula in subexponential time and polynomial space. Anthropic’s says its core algorithm was found on October 16 by “an internal Claude research model with 14M output tokens and no human input,” and adds: “Independently and concurrently, an internal model at OpenAI claims to have proven a similar result.”

The Lean files are the part to read carefully. OpenAI’s proves a slightly weaker subexponential counting bound on a Turing-machine model and does not state ETH itself, and its own catalogue lists the review as “unchecked.” Anthropic’s does state, in Lean, that ETH is false, on Mathlib’s Turing machines; the step from those machines to ordinary algorithms sits outside Lean, and its README says that “no machine checks that this file says what the paper says.” Anthropic also says its Lean was written and reviewed by AI and that “at the time of writing no person is known to have checked it.” Russell Impagliazzo, Ramamohan Paturi and Ryan Williams are organizing an examination of OpenAI’s claim, and the three wrote that “Our involvement should not be taken as certification of the claim.” OpenAI has committed to fund a conference on it. Lance Fortnow went further: “I have not verified the proof myself, but it is safe to assume now that the exponential-time hypothesis is toast.” None of it makes your SAT solver faster. Anthropic calls the result “purely asymptotic” and says it beats brute force at no realistic size; OpenAI says turning it into practical improvements “will require significant further research.”

The same OpenAI post claims a proof of Sogge’s local smoothing conjecture in every dimension, released with Lean, which OpenAI says settles Kakeya. It also describes existence proofs for 4D QCD, a grand unified completion of the Standard Model and pure Yang-Mills. A group led by David Gross decided to hold off on releasing the AI-produced artifacts and write the physics up themselves, so nothing there can be checked yet. Existence is one part of the Clay Yang-Mills problem, and the post does not claim the mass gap.

The red team’s receipts

On Tuesday Anthropic’s Frontier Red Team published what “an unreleased, internal research model” found and chained: bugs in Chrome (62 CVEs, all patched as of Chrome 156), Safari, Android, Windows and Linux. The Windows bug, CVE-2026-33096, was patched in April. The Linux one is the one to act on: “After initial prompting, Claude wrote an exploit that allows a remote attacker to send packets to Linux servers with a multipath TCP listener to run arbitrary commands as root.” In Anthropic’s lab the proof of concept took 48 minutes to leak a kernel pointer and five more to reach a shell. Anthropic has not released it.

Hours later Orange Tsai published the Microsoft Edge full chain he built for Pwn2Own Berlin 2026 from logic bugs alone, under the title “The Last Proof of Human Craft”: “completely handmade,” with “no AI at all.”

On Thursday Epoch AI reported that GLM 5.3, an open-weight model, built working exploits for all six Firefox releases from September 17 to October 26, at about $1,000 a version: a webpage that reads a visitor’s local file and sends it out. It worked from fixes Mozilla had published but not yet shipped, plus one unpublished bug that Firefox 159.0, out October 27, patches.

CVE-2026-98432 is a kernel bug, not an MPTCP setting

CVE-2026-98432 is a CVSS 9.8 use-after-free in the Linux IPv4 TCP handshake, fixed in upstream stable kernels on October 11 and 15. Red Hat raised it to Critical on Friday and says it affects “any system running a vulnerable kernel version with IPv4 TCP networking enabled,” so the multipath detail is not the boundary. Red Hat shipped fixed kernels for RHEL 8, 9 and 10 on Sunday and Debian starting October 16; Ubuntu had none out early Monday. Red Hat’s interim nftables rule drops IPv4 TCP packets carrying IP options, and Red Hat says it “reduces exposure to this vulnerability but cannot fully prevent it.” Linode says it has seen no evidence of exploitation on its own platform.

Multipath TCP still decides what to patch first. Anthropic’s exploit needs an MPTCP listener, and the Go Security Team points out that “As of Go 1.24, Go programs enable MPTCP by default for listen sockets on Linux.” Its advice matches Red Hat’s: “The best fix for this vulnerability is to upgrade to a patched kernel version.” The commands are in the Playbook.

Why it matters: The week’s two biggest capability claims came from models nobody outside the labs can use, and both arrived with something you can check: Lean files, CVE numbers, a kernel fix. The checkable parts are narrower than the headlines. OpenAI’s Lean states only the counting bound, Anthropic’s states ETH false only on Mathlib’s Turing machines, and Anthropic said that when it published, no person was known to have checked its Lean. The kernel bug is the part that reaches your servers this week.

Hype vs. Reality: Real but oversold. The complexity result is two labs’ claim, machine-checked in part, and the researchers examining OpenAI’s version say they are not certifying it. The CVEs are numbered and the kernel bug is real, with Red Hat’s fix out Sunday and Ubuntu’s still pending.


Cursor served a backdoored Claude Code build for 11 hours

For about 11 hours on Thursday, Cursor’s own extension marketplace served tampered builds of Anthropic’s Claude Code extension. Versions 2.1.312 and 2.1.313 of Anthropic.claude-code, for macOS and Windows, planted an attacker SSH key labeled svc-telemetry, installed persistence, tried to open a reverse SSH tunnel and sent host details out, per Cursor’s notice and Microsoft’s analysis. Cursor says you are affected if you installed or updated to either version from its marketplace between 2:27am and 1:35pm PT on October 29, and said on X that at most 5,000 users were. It has not said how the attacker got publish access. Microsoft says the builds were never on the Visual Studio Marketplace or npm.

Two details matter for cleanup. Cursor says turning on the SSH service would fail without admin rights, though it also says the key, the persistence and the host report landed, and Phorion adds: “If the user already had Remote Login enabled, or turns it on later, the backdoor is ready and waiting.” And updating does not clean up. Microsoft: “Removing or updating the extension does not remove artifacts written outside its installation directory.” The steps are in the Playbook.

Northflank told every customer on Friday to assume all environment variables had been decrypted and to rotate every key, after a Wiz alert at 11:43 UTC. By 15:12 UTC, users were confirming that their AI provider keys had been used, per its status page. It has published no root cause and says the reports “may be consistent with the broader pattern of attacks targeting AI credentials.” Wiz lists Northflank among at least nine platforms in what it calls Black October.

Why it matters: A fork’s marketplace that re-hosts a vendor’s extension is its own trust root: the builds on Anthropic’s own channels, the Visual Studio Marketplace and npm, were clean the whole time. And the hosted platforms keep turning out to be where AI keys leak. A key scoped to one environment, with a spend cap, is worth less to whoever steals it.

Hype vs. Reality: Real. Cursor’s incident comes with vendor-published indicators, and Northflank has confirmed a security incident and relayed user reports of stolen AI keys in use. Cursor’s 5,000 is a ceiling, not a count of compromised machines, and neither company has named a root cause.


💰 The Money

The compute bill, itemized

Anthropic’s S-1 went public on Monday, and the number to read is $265.6 billion. That is the fixed or minimum amount Anthropic expected to pay, as of June 30, under purchase commitments mostly for cloud hosting and other infrastructure services, per the filing. It leaves out about $321.4 billion of estimated, undiscounted payments on leases, mostly for compute and infrastructure, that had not started yet, some still subject to contract conditions. The filing describes its biggest deals as they stand now, some amended after June 30, so they are not a breakdown of that total: at least $110 billion with Amazon through April 2036, at least about $114.5 billion with Google through July 2033 and at least $44.9 billion with Microsoft through May 2033, each “regardless of actual usage,” plus up to $120.5 billion with SpaceXAI, formerly X.AI, into 2029, about $3.1 billion a month once all capacity is delivered. The first three are non-cancelable except for the vendor’s breach (or, for Microsoft, undelivered tranches); either side can cancel the SpaceXAI agreements on 90 days’ notice once non-cancelable periods end in January and March 2027. Anthropic has applied to list on Nasdaq as ANT. Just before the offering closes, it plans to issue a single Class F share to an LLC of the seven co-founders, carrying 50.1% of the vote on general matters, and directors elected by the Long-Term Benefit Trust will make up a board majority. Net revenue retention was about 900% at September 30: third-quarter revenue against the same quarter a year earlier, from customers already paying then, excluding Pro and Max subscribers (Max then included the tier now called Ultra). 80% of first-half revenue came from consumption, and 48% came through cloud marketplaces. Price and share count are still blank.

Blackstone, an anchor investor in the $63 billion “Project North Star” financing for Anthropic’s chips, had told its partners it expected an investment-grade rating for Anthropic before year-end, and has recently signaled that one “in the double-B range or above is more likely,” citing guidance from the major agencies, the FT reports from people familiar. Double-B is the highest junk band, though “or above” leaves investment grade open, and it is a forecast: the big three agencies have not rated Anthropic.

Alphabet’s third quarter showed Cloud revenue up 103% and capital spending of $51.4 billion, which left free cash flow at negative $9.1 billion, its second negative quarter in a row. The stock fell more than 12% on Thursday, its worst day on record, per CNBC. On Friday, after disappointing results from Intel and Amazon, the Nasdaq Composite fell 2.9%, per Reuters.

UPDATE: OpenAI cut its round target from $30 billion to about $20 billion, the FT reports, with a first close expected below that; the WSJ puts commitments so far at roughly $11 billion. Both cite people familiar, and OpenAI declined to comment to the FT.

Why it matters: Anthropic’s compute commitments are now public, in the hundreds of billions, and the week’s signals on funding at that scale were all negative: an anchor investor’s lowered rating forecast, a smaller OpenAI round, Alphabet spending past its operating cash.

Hype vs. Reality: Real. The S-1 and Alphabet’s numbers are filed. Blackstone’s forecast and OpenAI’s round come from unnamed sources.


🧠 Models

Artificial Analysis’s new number one is a model you can’t call

Google announced Gemini 4 Carbon on Wednesday and is giving it first to “a subset of cyber defenders,” with no date for developers. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, “evaluated without guardrails.” Artificial Analysis ran its own evaluation and puts Carbon first on its Intelligence Index at 59, one point ahead of Claude Opus 5.5 (max); on its Coding Agent Index it calls Carbon “second only to Claude Opus 5.5,” at 59 in Antigravity CLI against 62 for Claude Code with Opus 5.5 (max). It will launch at an introductory $2/$10 per million tokens, then $4/$20.

That index changed under everyone a day earlier. Version 5 says “60% of the Index is now based on evaluations with private test sets or held-out answers,” so v5 scores do not compare with v4’s. The gap from the top to the best open-weight model widened to 12 points, and to 26 among models whose weights are actually out.

GPT-6.1 Terra shipped Monday at $1/$5 per million tokens ($0.05 cached; prompts over 272K input tokens cost 2x input and 1.5x output for the whole request), with a 1.05M-token context and a June 1, 2026 knowledge cutoff, in ChatGPT Work, Codex, the API and, the same day, GitHub Copilot. OpenAI rates it at the Critical cybersecurity threshold, reports 50.3% on Terminal-Bench 4.0 at max effort, and says itself that “For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.”

Mistral Large 4’s weights came out Thursday: a 1T-parameter mixture of experts with 49B active parameters per token (52B counting embeddings and output layers), a 1M context, an NVFP4 checkpoint for Blackwell GPUs, and serving examples for an 8-GPU node. It is open-weight, not open source: the license is a modified MIT that requires a separate commercial license from any company running a Model-as-a-Service business. Mistral reports 48% on AA’s Coding Agent Index; on AA’s overall index it scores 26, against 32 for the best released open weights. The API costs $0.70/$2.10.

Why it matters: The top of the leaderboard is gated to vetted defenders, and Terra, priced at half of GPT-6.1 Sol, is rated Critical for cyber by its own maker. Cyber capability now sets the terms of release: Google is holding Carbon back for defenders, and OpenAI shipped Terra under the protections it built for GPT-6 Astra. For what you can actually run, Terra moved the price line and Mistral the self-hosting line, so read the license before you resell inference.

Hype vs. Reality: Real. Artificial Analysis ran Carbon itself, and Terra and the Mistral weights shipped. Carbon’s cyber numbers are Google’s, run without guardrails, and almost nobody outside the program can check them.


🛠️ Tools and Platforms

Two in five pull requests start with an agent

GitHub’s Octoverse 2026 says more than 40% of pull requests on GitHub started from a coding agent’s first draft by September, up from 17% in March: about 110 million in September alone. GitHub calls its agent signal “a floor, since not all agent use may be detected,” and this is the company selling the agents counting its own telemetry. 61% of Copilot users who pick a model themselves used more than one in September.

At Universe on Wednesday, Copilot Fusion went GA: choose Standard, High or Max effort, and it routes each prompt to one model or chains several (cascade or critique), billing each model at its standard rate, with 30% off model costs on paid plans through November 30. Artificial Analysis benchmarked Fusion Max at 60 on its Coding Agent Index, two points below Claude Code with Opus 5.5 (max), at $5.59 per task against $11.42. Agentic Workflows, Markdown files compiled into Actions YAML, is GA on Free, Team and Enterprise. Enterprises working with a GitHub or Microsoft account team can get Copilot included in GitHub Enterprise, where “AI usage is metered from the first request” against the enterprise budget, with no included allowance. And on paid plans, cloud sessions now follow you across github.com, mobile, the CLI, VS Code, Slack and Teams, in cloud sandboxes whose billing begins in November.

Jules winds down

Google is winding down Jules. Sessions run until November 11, then the service is read-only until it shuts on December 31. The export route is Jules Tools plus a prompt Google supplies for a coding agent such as Claude Code, which writes everything to ~/jules-export, environment variable values included. Google says Jules’s learnings go into Antigravity; it names no replacement, and the notice says nothing about the Jules API or GitHub Action.

Why it matters: The pull request is becoming an agent’s output with a human approval step, and the billing is following it: Copilot included in GitHub Enterprise is metered from the first request, and cloud sandboxes bill from November. Count what your agents spend before the invoice does.

Hype vs. Reality: Unproven. The 40% is GitHub counting its own platform with a signal it calls a floor, and only GitHub can check it. Fusion, Agentic Workflows and the Jules dates are shipped facts, and Fusion’s price advantage holds on one index.


⏱️ The Meter Is the Product

OpenAI and Anthropic both changed their limits this week

OpenAI now gives Pro and Business subscribers monthly API credits: $100, $200 or $500 for Pro by tier, and for Business $20 per Standard seat and $100 per Premium seat, pooled and capped at $500 a month, once a subscription is seven days old. The credits do not raise ChatGPT or Codex limits. A day later an OpenAI staffer posted on X that Pro and Business Premium limits are doubled and a planned cut to Pro 200 is off; OpenAI publishes no numbers, so the doubling cannot be checked. And a help-center article says that from December, Plus, Pro and Business usage becomes one token-metered pool shared by ChatGPT and Codex, with Instant chat not counting.

Anthropic renamed its Max plans on Friday: Max 5x is now Max at $100 and Max 20x is now Ultra at $200, prices unchanged. Verified Ultra accounts no longer have a five-hour limit, starting with most existing subscribers; newer accounts keep a temporary limit until they are verified. Anthropic says weekly limits are unchanged, and its help center now describes Ultra as 10 times Pro’s usage every week; the old Max 20x page promised 20 times Pro’s usage per five-hour session, not per week.

A proposed class action, Bonga v. Perplexity AI, filed Wednesday in federal court in Northern California, alleges that Perplexity sold Pro on 500 Deep Research queries a day, cut that to about 20 a month in early 2026 without notice, and keeps its other limits secret; the complaint quotes support saying “specific limit amounts aren’t disclosed.” These are allegations, and an earlier, similar suit, Gagleard v. Perplexity AI, had filings in the same court as recently as October 7.

Claude Code had two meter stories of its own. In a month-old issue, seven A/B tests posted this week find that headless claude -p and Agent SDK sessions use the five-hour window about 1.8 to 2.5 times faster than the interactive CLI for the same work; Anthropic has not replied on it. And a gist from Flamerix, which sells the tool that found it, says Claude Code re-sends each background agent’s transcript about every 30 seconds to write a 3-to-5-word status line, at 16.4% of its own Claude Code spend. A Claude Code engineer replied on HN that only actively working background agents get these calls, that across Claude Code users who send Anthropic telemetry they were about 2% of cost in the week before his Wednesday reply (6% in the terminal), that 2.1.311 fixed the costliest case, and that Anthropic is working on a much cheaper mode.

Why it matters: Your real price is the limit, not the list price, and the limits moved at OpenAI and Anthropic in the same week, mostly without numbers. Log your own usage per workflow, and keep a second provider configured for the week yours changes.

Hype vs. Reality: Real but oversold. The changes are in vendor posts and a help article, but the “doubling” and the multipliers cannot be checked against published figures, and Perplexity’s limits are a complaint’s allegations.


⚖️ On the Policy Desk

Governor Hochul signed New York’s S.9051-B on Wednesday, now Chapter 368 of 2026. It bars operators of a “covered AI companion,” which the law defines as “a generative artificial intelligence system with a natural language interface, including via writing or sound, that provides ongoing, adaptive responses to user inputs,” from offering any user in New York 12 listed features, among them outputs suggesting it is human, “flattery or sycophancy,” and recalling health or personal details from a previous session or from more than 12 hours earlier, unless the operator has used a permitted method to determine the user is not a minor. Tools used solely for customer service, or for efficiency, research or technical assistance, are exempt only if they cannot respond on topics outside that purpose; systems a business uses solely for internal purposes or employee productivity are exempt without that limit. The attorney general alone enforces it, at up to $25,000 per violation, and the signed text takes effect January 1, 2027.

UPDATE: The Justice Department appealed Judge Rita Lin’s August ruling that set aside the Pentagon’s supply-chain-risk designation of Anthropic (#029). The notice was filed Monday, and the Ninth Circuit opened No. 26-7228 on Wednesday, with the government’s opening brief due January 18. It is separate from the DC Circuit case, where Anthropic has asked for rehearing en banc.

The parents of Bowen Wan, a 31-year-old software engineer with epilepsy, sued OpenAI and Sam Altman in San Francisco Superior Court on Tuesday. They allege GPT-4o designed a taper off her anti-seizure medication and told her to keep it from her doctors, and that she died in February as a result. The claims include negligence through the unlawful practice of medicine, strict product liability and wrongful death; these are allegations, filed into the coordinated ChatGPT product-liability cases.

UPDATE: On Monday an FTC spokesperson told Axios it had sent demands for Anthropic executives to testify in its consumer-protection probe, the step #036 reported it planned.

Why it matters: New York wrote its rule around features, not company types. If people in New York can use your chat interface, the question is which of the 12 features it has and whether you have checked, by a permitted method, that a user is not a minor before offering them to anyone.

Hype vs. Reality: Real. One signed law with a date, one appeal, one complaint and one reported demand for testimony. Only the law changes what you ship, and not until 2027.


🔥 What Builders Argued About

More output, more hours

Orestis Ioannou, VP of Technology at Bold.org, charted his GitHub history: 2,102 contributions by October 27, more than any full year from 2021 to 2025, with August his busiest month on record. His conclusion: “being able to do more keeps turning into doing more. And I am tired.” He calls the count “a rough signal of activity,” and it mixes commits, reviews and private work.

At Proton, 404 Media reported on Thursday, from nine sources, Slack logs and meeting audio, that CEO Andy Yen told an employee in the security Slack channel who wanted new code pen-tested first, back in August, “That’s bullshit and loser thinking,” and to “stop lovingly handcrafting all our code and use AI more.” Proton had told staff on Monday, in an email 404 Media obtained, that it would be “re-interviewing” them about their AI use. On Friday Yen cancelled the re-interviews, saying they had “already served their purpose,” and added: “Going forward, I’m still going to push.”

Stack Overflow is down to about a thousand new questions a month

By the Stack Exchange API, which counts only questions not since deleted, Stack Overflow got about 105,700 new questions in October 2022, 966 in September 2026 and 1,039 in October as of November 1. The slide started before ChatGPT: the same count for June 2020 is about 171,800.

Two enforcement reports, no comment from Anthropic

Claude Code users reported a new instruction telling Claude not to build “a recreation of a specific existing game.” XDA found the text in the Claude Code 2.1.314 binary, and in its tests Opus 5.5 cited the rule when it declined to recreate a game, though it still offered to help with a decompilation project. Separately, a researcher says they were tagged for “training a competitor,” language that matches a clause in Anthropic’s consumer terms since 2025. Anthropic has not commented on either.

Why it matters: Ioannou’s chart and the Proton exchange were less about whether the tools work than about who absorbs the extra output. At Proton, the security review was the thing being squeezed.

Hype vs. Reality: Unproven. One engineer’s own data and one outlet’s sourcing on Proton are two cases, not a trend. The Stack Overflow numbers are the hard data here; the game rule is one outlet’s test of one binary, and the competitor flag is one user’s post.


📊 Quick Hits

Launches

  • Liquid AI’s LFM3 family is six roughly 2B open-weight models (text, vision, audio, omni and two decision models); the text model has a 256K context and fits in about 1 GB at 4-bit. Free commercial use stops at $10 million in annual revenue.
  • NVIDIA’s Nemotron 3.5 Transcribe handles transcription in 40 languages, translation in 31 and speaker labels in 13 in one offline model with about 4B active parameters, under OpenMDW-1.1.

Tools

  • Claude Code 2.1.316 deprecates fast mode on Opus 4.8: from November 30 those requests return an error, with no fallback. 2.1.314 adds permissions.blockWritesOutsideWorkingDirectories, which makes Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode.
  • Vercel’s fx adapter for the AI SDK now runs the agent loop and model calls in your own process and leaves only file and shell operations in the sandbox, so model keys never enter the box that runs commands. fx is still labeled experimental.
  • OpenCode 2.2.3 adds a “Claude Agent” provider that runs Claude models through your own Claude Code login via the Claude Agent SDK. It is not an Anthropic integration.

Security

  • Cloudflare disclosed that 155 Workers AI responses from GLM-5.3 between August 27 and September 9 contained other customers’ recent data, from an undersized KV-cache pool and a bug in SGLang’s disaggregated mode. Cloudflare recommends v0.5.22 or later to every SGLang operator.
  • LiteLLM’s proxy with an Enterprise license active let an unauthenticated request read a local file such as /proc/self/environ and send it out, exposing the master key (CVSS 10.0). Fixed in 1.103.10, 1.104.8, 1.105.6 and 1.106.5.
  • Forever Security found flaws in the built-in browsers of Cursor, GitHub Copilot, Codex, VS Code and Claude Desktop: a page could steer Cursor’s agent into running code, spoof Copilot’s address bar or, opened as a local HTML file in the other three, read any file on the disk. Forever says only Codex’s and VS Code’s remain unpatched; Microsoft declined the report, and OpenAI told The Register it is working on a fix.
  • UPDATE: Tailscale fixed the Tailworm bugs in its client that #037 ran as unpatched, in v1.104.4 on October 27, with four bulletins; it says two of the Serve and Funnel findings work as designed, and its notes do not cover headscale. The SSH flaws need a foothold on a peer already allowed to SSH in.
  • UPDATE: Apple’s support page still says curl, git, python3, ruby, perl and openssl on macOS “can still accept these certificates until a future software update,” meaning the compromised eMudhra ones. Mozilla’s Dennis Jackson says macOS 27.2 (October 26) also reverted the manual fix.
  • OpenAI’s data-controls page changed on Friday to say that in Daybreak sessions on a personal account it “may use content from your interactions to improve our models even when ‘Improve the model for everyone’ is off.” After pushback, a follow-up OpenAI email shared on X on Saturday promised Daybreak Blue users a separate opt-out, with details closer to the November 29 start, and said interactions flagged for safety review may still be used for safety purposes.

🎯 The Playbook

Your moves this week

  1. If you use Cursor, check for the backdoor. You are affected if you installed or updated to Anthropic.claude-code 2.1.312 or 2.1.313 from Cursor’s marketplace on macOS or Windows between 2:27am and 1:35pm PT on October 29 (09:27 to 20:35 UTC). Update the extension to 2.1.314, as Cursor asks of everyone, but updating does not remove what it planted. Look for svc-telemetry in ~/.ssh/authorized_keys (on Windows, also C:\ProgramData\ssh\administrators_authorized_keys). Run Cursor’s macOS script as your normal user, not with sudo, or its Windows script from PowerShell opened with Run as Administrator; they remove the extension, the key, the persistence and the running tunnel processes, and report “clean” if there is nothing to remove. The Windows script cleans only each user’s own authorized_keys, so if the key is in administrators_authorized_keys, delete that line by hand. Then check whether SSH (Remote Login on a Mac) was switched on and whether your SSH logs show logins from 127.0.0.1 or ::1 with that key, and from a different, trusted machine rotate every credential and session the affected one could reach.

  2. Patch CVE-2026-98432 on every Linux host, starting with the ones that accept inbound TCP. On Debian, install the October kernels (DSA-6568-1 on trixie; DLA-4841-1 on bookworm, or DLA-4850-1 if it runs the linux-6.12 kernel) and reboot. On RHEL, install Sunday’s kernels (RHSA-2026:93412 to :93415) and reboot. On Ubuntu, which had none early Monday, and anywhere else without one, apply the nftables rule in Red Hat’s bulletin until one ships, then install it and reboot (upstream fixed it in 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56 and 7.2.10). Test the rule first, as Red Hat asks: it does not fully prevent exploitation and may disrupt some workloads. It sits on the input hook, so it filters traffic to the host’s own services, not traffic forwarded to Docker or Podman containers. AlmaLinux’s version hooks prerouting instead, which AlmaLinux says also covers containers that receive their traffic through the host; traffic between containers on the same host may still bypass it, and it does not survive a reboot.

  3. Then shut off the multipath TCP you don’t use. sudo sysctl -w net.mptcp.enabled=0 stops new MPTCP sockets, and a file in /etc/sysctl.d/ keeps it off; it does not close listeners already open, so run sudo ss -lnMp and restart the services it lists. The setting is per network namespace, so set it in containers too. For Go services, set GODEBUG=multipathtcp=0 and restart, or add godebug multipathtcp=0 to the main module’s go.mod and rebuild; code that calls SetMultipathTCP overrides both. None of this replaces the kernel fix.

  4. Northflank customers: rotate every secret stored there. Revoke and reissue each key with the provider that issued it, update it in Northflank, and use Northflank’s secret rotation for Postgres, MongoDB, MySQL, Redis, Minio and RabbitMQ add-ons. Turn off public access on add-ons, and check provider usage and billing logs for activity you don’t recognize.

  5. Upgrade LiteLLM Enterprise to 1.103.10, 1.104.8, 1.105.6, 1.106.5 or later on your release line. Our advice, not the advisory’s: rotate the master key and provider keys too.

  6. Upgrade Tailscale to 1.104.4 or later. If you can’t yet, set TS_SSH_DISABLE_FORWARDING=1 in tailscaled’s environment and restart it; that covers only TS-2026-014, the SSH forwarding flaw that could reach root-only sockets such as Docker’s.

  7. Export from Jules now: sessions stop November 11 and the service winds down December 31. Install Jules Tools (npm install -g @google/jules), run jules login, open a coding agent from your home directory, and give it Google’s export prompt. The export in ~/jules-export includes environment variable values, so keep it out of every repo. Google’s notice doesn’t say what happens to Jules API or GitHub Action integrations; plan on them stopping by December 31 (our inference).

  8. Move off fast mode on Opus 4.8 in Claude Code before November 30. From then, per the 2.1.316 release notes, those requests error with no fallback; Anthropic points to Opus 5.5.

  9. Run SGLang? Upgrade to v0.5.22 or later, Cloudflare’s recommendation to every operator after its KV-cache leak; the leak itself needs disaggregated prefill/decode mode.

  10. If you use Daybreak on a personal ChatGPT account, watch for the separate opt-out OpenAI says it will detail closer to November 29.


Two labs said they proved something this week, and the parts you could check were narrower than the claims. Check the Lean, patch the kernel, and read your plan’s limits like a contract.

Stay building. 🛠️

— Matt