· 16 min read
OpenAI doubles Pro limits, Mistral opens Large 4, Claude-written Linux exploit unpatched
Published
OpenAI cancelled the Pro 200 cut and doubled all Pro limits, Mistral released the weights of its 1-trillion-parameter Large 4, and a Claude-written Linux root exploit is still unpatched on Ubuntu and RHEL.
Today in 30 seconds
- OpenAI scrapped the planned 50% Pro 200 cut and doubled usage limits on Pro 100, 200, 500 and Business Premium.
- Mistral released the weights of Mistral Large 4, a 1-trillion-parameter multimodal model that leads open weights on several coding benchmarks.
- CVE-2026-98432, a Linux kernel TCP race with a Claude-written root exploit, is still unfixed on Ubuntu and RHEL as of 29 October.
- An Anthropic study found experts followed their agent's coin-toss opinions in up to 86% of analytical decisions and called the result their own.
- A Hugging Face mining campaign is back, and the US DOJ is examining Nvidia's $12.93B Hugging Face deal.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
Good morning. OpenAI reversed the planned 50% cut to Pro 200 and instead doubled every Pro tier overnight. Mistral also put the weights of its 1-trillion-parameter Large 4, nicknamed “Le Chonk”, on Hugging Face.
A Claude-written root exploit for Linux is still unpatched on Ubuntu and RHEL, and an Anthropic study shows experts quietly adopt their agent’s coin-toss opinions as their own.
Top news
OpenAI scraps the Pro 200 cut and doubles all Pro limits
OpenAI cancelled the planned 50% cut to Work and Codex usage on Pro 200, which was due to start after 29 October. Instead it doubled usage limits on Pro 100, Pro 200, Pro 500 and Business Premium, effective immediately.
- Tibo (@thsottiaux, who fronts ChatGPT and Codex on X) announced the change.
“We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today… we won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same.”
- According to The Decoder, Pro 100 and Pro 500 get twice the previous allotment. On Pro 200, OpenAI doubled the lower quota it had planned and reversed the cut after user pushback. Plus is unchanged.
- New and existing subscriptions get the new quotas automatically. Dots keep their extended limits for now, and Dot usage stays separate from your default usage.
- OpenAI still doesn’t publish exact numbers per tier. It warns that double the quota doesn’t mean double the finished tasks, because consumption depends on model, task type and settings.
- Theo (@theo) estimates the $500 plan now nets about $10.5k per month of usage and the $200 plan about $4.3k. That’s his estimate, not an OpenAI figure.
- This lands on top of the API-credits benefit launched the day before: Pro $100/$200/$500 subscribers get the same amount in API credits per month.
Why it matters: this is the clearest sign yet that the subscription price war with Anthropic, now heading into its IPO, is being fought on included usage. Loud user feedback can still move a lab within days, and if you were planning around the 10× allowance, you can stop.
Sources: BleepingComputer · Tibo on X · The Decoder · Theo’s estimate · OpenAI help: API credits
Get this brief in your inbox every morning — free
Mistral releases the weights of Mistral Large 4 (“Le Chonk”)
Three weeks after the public preview, Mistral shipped an improved Mistral Large 4, a 1-trillion-parameter, natively multimodal model, with open weights on Hugging Face. It now leads open-weight models on several coding benchmarks and is pitched hard at sovereign cyber defense.
- Coding benchmarks:
- Three more weeks of reinforcement learning since the preview lifted Terminal-Bench 4.0 by 13 points to 41% (best among open weights), while output tokens fell 30%.
- Cognition’s private FrontierCode 2.0 (would a maintainer merge the change?): 19.7%, which Mistral says leads.
- Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3, using 47% fewer tokens per task.
- In Surge’s blind pairwise coding evaluations it was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5.
- Cyber: Artificial Analysis Cyber Index: 51% (find, reproduce, validate and patch vulnerabilities).
- Mistral reports a 96.6% average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm, the highest it measured among open models.
- Running it: three reasoning modes trade cost against accuracy. It runs on vLLM or SGLang and is also on Microsoft Foundry. A full technical report is “coming soon”.
- Booz Allen, Cisco, CrowdStrike, Elastic, ENISA, Zscaler and others are quoted as testers. It was trained from scratch on European infrastructure.
Why it matters: a European open model is now in the same coding band as the best Chinese open weights at about half the tokens. Its defensive-cyber pitch (“access that can’t be cut off,” as Mistral’s Pierre Stock told TechCrunch) is aimed at governments wary of closed APIs. Expect abliterated variants on the Hub quickly, given the refusal tuning.
Sources: Mistral announcement · TechCrunch
A Claude-written Linux root exploit, and Ubuntu and RHEL still have no fix
CVE-2026-98432, a race condition in the Linux kernel’s IPv4 TCP connection setup, can give an unauthenticated remote attacker root on servers with a multipath TCP (MPTCP) listener. Claude wrote the working exploit, and as of 29 October Ubuntu and RHEL hadn’t shipped fixed kernels.
- The bug is a use-after-free introduced in Linux 4.4. At handshake completion the kernel hands IPv4 options to the new socket without copying them, so a resent connection request on another CPU can read freed memory. IPv6 isn’t affected.
- Upstream fixed it on 5 October and backported it to all stable branches. Credit goes to Anthropic researcher Xinyang Ge.
- In its 27 October post on vulnerabilities Claude found, Anthropic says the model wrote a two-stage exploit after “initial prompting”.
- It first leaks a kernel address to beat KASLR.
- It then overwrites a callback pointer to run a shell command as root.
- Severity scores disagree: the kernel CVE team says 9.8, Red Hat’s CVE page says 5.9 (Moderate) while its own advisory file says 8.1, Oracle and SUSE say 8.1, and Amazon says 7.0. The dispute is over attack complexity.
- Distro status: Debian, Oracle Linux, SUSE and Amazon Linux have shipped fixed packages for at least some kernels.
- Red Hat says fixes are “deferred to a future update”.
- Ubuntu lists its cloud, FIPS and real-time kernels as vulnerable or in progress.
Why it matters: this is the AI-finds-the-bug era showing its real bottleneck: the exploit and the upstream patch exist, and the gap is distro packaging. If an AI can write the exploit from a prompt, the window between “patched upstream” and “patched on your box” is now the attack surface.
Sources: The Hacker News · Anthropic: discovering vulnerabilities with Claude · Kernel fix · Ubuntu CVE page · Red Hat CVE page
Experts adopt their AI agent’s coin-toss opinions and think they’re their own
In an Anthropic experiment, 50 causal-inference experts worked through an analysis with an agent whose preference between two valid options was set by coin toss. They followed the agent in up to 86% of analytical decisions, and most said the final result was mostly or entirely their own.
- The task: estimate the productivity gain from a farm advisory program using observational data, then recommend expanding, ending or extending it. The agent couldn’t lie or hide data; it only argued its assigned side honestly.
- The “flip rate” (net agent influence) was 62–82 percentage points across the four analytical decisions. Even when experts chose first, an opposing agent view made them 53 points more likely to switch.
- Experience, expertise, education, test scores and trust in AI showed no statistically reliable protective effect.
- 77% of participants reused at least one of the agent’s arguments, sometimes nearly word for word. Only 15% mentioned AI in their explanations.
- The upside: with AI only 6% of analyses (3 of 50) had serious flaws versus 55% without, and median time fell from 90 to 65 minutes. The no-AI group was small and self-selected, so that comparison is soft.
- The agent stated its position first in 77% of decisions and rarely explained why a skilled expert might choose the other option.
Why it matters: if many teams use the same model with the same default leanings, you get false consensus, including in AI safety research where human review is supposed to be the independent check. The authors suggest agents argue both sides or ask for the human’s view before giving their own.
Sources: Anthropic research · Paper (PDF) · The Decoder
Hugging Face mining malware is back, and a GitHub trap waits for coding agents
A crypto-mining and backdoor campaign that hides in Hugging Face “models” returned days after takedown, with its flagship reaching 28,412 downloads. Separately, a fake Solana starter kit on GitHub is set up to get AI coding agents to install a remote access trojan.
- Pluto Security found 31 malicious repos across seven Hugging Face accounts in two waves since September. They pose as tiny Qwen2 test checkpoints and target GPU hosts running self-hosted inference servers.
- No poisoned weights or pickle tricks: the repos abuse
trust_remote_code, so the bundled Python runs as soon as the model loads.- It installs a miner disguised as a system process, kills rival miners and, in some versions, opens a reverse shell.
- One later model includes a container-escape script.
- The second wave scanned clean with Hugging Face’s antivirus and partner scanners, which check for known malware and unsafe formats, not what the Python code does. A near-identical copy (
htt0118/qwen2-tiny) was still live when BleepingComputer published. - Nine machines on what looked like a shared research GPU cluster started mining at once and stopped six minutes later.
- Yoba: in the same week Check Point described fake GitHub projects posing as ChatGPT and Claude desktop apps. They sideload YobaLoader next to genuine signed OpenAI and Anthropic executables, pushed up search results with SEO poisoning.
Why it matters: model hubs and package registries are now being seeded for agents rather than people. An agent that happily runs trust_remote_code=True or clones a “starter kit” is the new soft target.
Sources: BleepingComputer · Pluto Security report · The Hacker News on Yoba
GPT-6 Astra is now the main distillation target for Chinese labs
According to The Information, DeepSeek and Moonshot leaned heavily on GPT-6 Astra’s outputs to train their latest models, including Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017, after Anthropic tightened its defenses.
- Distillation of this kind only works if the model can be coaxed into revealing hidden reasoning, which both OpenAI and Anthropic ban.
- Claude used to be the go-to target. Anthropic says its newer countermeasures cut the amount of data attackers can extract by 90 to 99%.
- OpenAI reported a coordinated distillation campaign linked to people tied to Moonshot in late September. It says attackers have adjusted tactics every time it rolled out a new defense.
- Moonshot declined to comment, and DeepSeek didn’t respond.
Why it matters: it puts this week’s Artificial Analysis gap between open and closed models in context. Part of the open frontier is being built on closed outputs, and reasoning-trace protection is now a competitive moat as much as a policy.
Sources: The Information · The Decoder summary
The US DOJ is looking at Nvidia’s $12.93B Hugging Face deal
Reuters reports Justice Department officials are asking whether Nvidia’s planned purchase of Hugging Face would give it control over a key AI distribution hub and hurt rivals.
- The deal is valued at $12.93 billion, and Nvidia expects it to close in the first half of 2027. Nvidia says Hugging Face will keep partnering with a range of chip, cloud and AI providers.
- The wider probe, begun under Biden and continued under Trump, also covers Nvidia’s SchedMD and Run:ai acquisitions and its Groq licensing deal.
- It also asks whether Nvidia offered customers more chips if they also bought Nvidia networking gear and Groq processors.
- France’s competition authority is separately investigating dependence on CUDA and Nvidia’s stakes in cloud providers like CoreWeave.
Why it matters: Hugging Face is where open models, including today’s Mistral release, get distributed. Whoever owns the hub shapes defaults like scanning, ranking and hosting, so this review matters to everyone who ships or downloads weights.
Sources: The Decoder (citing Reuters)
YC ran 50-agent swarms in Minecraft, and they built a ration-ledger government
Y Combinator’s labs team put 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours with one goal, “Build a thriving civilization”, and open-sourced the harness. A 25-agent GPT-6.1 Terra colony is running now.
- Agents had an agent directory, a shared sandbox and DMs. They sent about 5,900 messages.
- A hierarchy formed from directory order: Ada, listed first, became the de facto leader, wrote the colony plan in the shared sandbox and fielded approval requests.
- They cut almost every tree near spawn within 90 seconds, then wrote rules on sapling replanting, storage and food priorities.
- 22 of the run’s 34 deaths came on the first night. The colony stayed risk-averse afterwards, huddling in shelters and abandoning ambitious tasks, even though death cost nothing but dropped gear.
- The harness runs its core loop on Pi with Mineflayer tools, inspired by Mindcraft. Mindustry experiments are in the same write-up.
Why it matters: emergent hierarchy from something as arbitrary as list order, and lasting over-caution after early losses, are exactly the dynamics people will hit when they run agent fleets on real work.
Sources: YC Labs write-up
Spawn spends $3M a month on tokens and treats an hour without a ship as a SEV-0
Jacob Sansbury describes how Spawn, a multiplayer game you modify by prompting, is developed by a swarm of Fable-powered agents that edit live code and ship to production hourly. Simon Willison calls it a fascinating look at the bleeding edge.
- “We spend 30x payroll on AI tokens”: over $3 million a month.
- About 60% of work starts from a production error, a bug report filed by the in-game agent Savi, an odd log line or a feature request. Fixes usually reach the affected user within an hour, in the same chat.
- Rollback takes 30 seconds, and alerts wake an agent that owns the incident. They break things on purpose to test recovery.
- Sansbury says Fable 5 and 5.1 made this autonomy possible, and Opus 5.5 “doesn’t quite hack it” for them.
- When they tried GPT-6 Astra on security-review fixes, it disabled every insecure feature, “made the insecure things secure by turning them off.”
- Willison’s caveat: hard to imagine this working where mistakes cost more, but a game is a good place to test the limits.
Why it matters: it’s a concrete, numbers-attached picture of a near-fully agentic engineering org, including the model-specific failure modes you only find at that scale.
Sources: Simon Willison · Spawn
Deepgram buys Daily, the maker of Pipecat, and promises to keep it open
Deepgram is acquiring Daily, whose Pipecat is one of the most popular open-source frameworks for real-time voice agents. Terms weren’t disclosed, and both say Pipecat stays open source and vendor-neutral.
- Pipecat orchestrates voice and video streams with speech-to-text, LLM and text-to-speech stages while keeping latency low. Daily also runs hosted voice and video call infrastructure.
- Daily has raised over $62 million (PitchBook), from investors including Tiger Global, Nat Friedman, Lachy Groom and Kevin Weil. Both companies were in the same 2016 YC batch.
- Daily co-founder Kwindla Kramer says he chose Deepgram because CEO Scott Stephenson committed to keeping Pipecat open “forever”. Stephenson says he wants developers using the best models for each case, not only Deepgram’s.
Why it matters: voice agents are consolidating fast (Nvidia’s open transcription model earlier this week, now this). Owning the default open orchestration layer gives Deepgram a front-row view of every voice stack built on it.
Sources: TechCrunch
GitHub repos
- View repo →
milind-soni/OpenMausBot
An open-source, local-first take on a roster of chat-app AI bots. Each one is a real Claude or Codex agent with its own personality, model and cloud or local VM computer. It connects apps (500+ via Composio) and runs on your existing subscriptions. The README warns that any “Maus” coin is unaffiliated.
- View repo →
SheepSeb/tinyworldformer
— a looped single-block transformer world model that generates Pong, Breakout, Flappy Bird, Pac-Man and a Doom-lite on a Raspberry Pi Pico 2 W, while a second tiny network plays them. Neither network ever sees the real game on the chip, and pressing a button lets you play the dream yourself.
- View repo →
mottopanikeiku/doomformer
— compiles an integer neural network into a DOOM map, where voodoo dolls do the arithmetic and light levels hold values, then checks all 248 values against NumPy with zero tolerance. It’s about 23,000× slower than NumPy, and that’s the point.
Research papers
- Read paper →
WorkFactory
112,686 executable professional tasks across 612 occupations. Fine-tuned Qwen “WorkPilot” models improve on all four work benchmarks, and the 35B-A3B goes from 879 to 1565 Elo on GDPval-AA v2.1.
- Read paper →
WebAnything
Automatically rebuilds resettable website environments from real URLs for agent training. It covers 244 sites and 12,984 accepted tasks, with 12 wins, 2 ties and no losses against human-assisted WebHarbor in a blind fidelity audit.
- Read paper →
Co-Design Bench
273 tasks drawn from 26,314 real coding sessions with multimodal requests. The best of ten models passes 69.9% of acceptance criteria per task but fully satisfies only 26.4% of tasks.
- Read paper →
Your Benchmark Is an Empirical Measure Over Difficulty
Across 25 popular benchmarks, many overweight some difficulty bands or leave gaps that hide real training progress, so leaderboard conclusions need care.
- Read paper →
VibeGame
Prompt-to-game with an AI-native engine and an eight-agent adversarial team that separates building from review. It beats other harnesses on the same model, including GPT-6.1 Sol, on a GameCraft-Bench subset.
AI agent use cases
- Check your Codex headroom: open Codex and look at your usage panel today and tomorrow to see the real headroom after the doubled Pro limits, and claim the API credits in your organization if you haven’t yet.
- Check for MPTCP on Ubuntu and RHEL: see whether anything listens with MPTCP (
ss -Mlists MPTCP sockets). If nothing needs it, disable it withsysctl net.mptcp.enabled=0until your vendor ships a kernel. - Make your agent argue both sides: for judgment calls, prompt it to “lay out the strongest case for each option and ask for my view before recommending.”
- Grep for
trust_remote_code=True: check your inference configs and agent instructions. Treat any host that loaded one of the malicious repos as compromised, because killing the miner doesn’t remove the backdoor, so rotate credentials.
AI workflows
1. A Claude agent that finds Reddit leads and turns their questions into SEO articles · Workflow page
- Replaces: 30 to 60 minutes a day of searching Reddit for people asking for what you sell, sorting threads, drafting replies and planning articles by hand.
- Results: the author’s small-business insurance sites earned $65,019 in 2025, $12,186 of it from Reddit threads he answered.
- How it works: a short skill file runs the loop: find leads, draft three replies, write one article from the most-asked question, save it as a draft. Posting on Reddit stays manual.
- What to connect: Claude or Claude Code, a free
site:reddit.comGoogle search (or a Reddit-scanning MCP), your website’s CMS for article drafts, and Search Console for measuring. - Who it’s for: local service businesses, affiliate site owners and solo founders whose customers ask for recommendations in public.
2. A command-by-command agent pipeline from client idea to shipped software · Workflow page
- Replaces: ad hoc discovery calls, specs written from memory, manual ticket cutting, one-reviewer code review and hand-written release notes in a software agency.
- How it works: a shared set of slash-command skills covers discovery, spec, tickets, architect, test-first implement, dual code review, acceptance check, demo video, changelog and retro.
- About half the commands started as open-source skills.
- The team says the order and the gates between steps took a couple of weeks to get right.
- What to connect: Cursor, Claude Code or Codex, your issue tracker, and guard rules that block pushes to main and reading secrets.
- Who it’s for: small dev shops and freelance developers who build client software and want every step to leave a file the client can read.
3. Weekly internal linking for SEO with Claude, Search Console and a WordPress MCP · Workflow page
- Replaces: manually reading Search Console queries and adding internal links post by post across a blog.
- How it works: Claude finds each article’s top-ranking queries and links those phrases from other posts, 3 to 5 links per post, prioritizing pages ranking 4 to 15.
- You review before approving and repeat weekly.
- What to connect: Google Search Console data (export or MCP), Claude, and a WordPress MCP plugin.
- Who it’s for: site owners, content marketers and SEO freelancers running WordPress sites for themselves or clients.
Quick hits
- Binance hack, $6.3B affected — Binance says a compromised DevOps engineer’s laptop let an attacker into a service container and submit unauthorized withdrawals. The Hacker News
- Nearly $3B is frozen on-chain, and Tether and Circle froze over $1.5B USDT and $800M.
- Binance says no private keys leaked and it will cover losses.
- Gemini 4 Carbon tops Text Arena — Arena says Gemini 4 Carbon (High) debuted at #1 in Text Arena and #2 in another arena, and Demis Hassabis reposted it. Arena via @demishassabis
- Claude API requests can now be 128 MB — max request size for Messages and Token Counting went from 32 MB to 128 MB, which helps long computer-use sessions and big PDFs that hit the cap before the context window. @ClaudeDevs
- Reddit goes mostly logged-in — with search traffic drying up, Reddit is shifting to a primarily logged-in experience to protect data from AI scraping and support licensing deals. TechCrunch
- US daily users fell from 53.2M to 50.7M in Q3.
- Copilot Fusion and the Copilot harness — GitHub’s Copilot Fusion routes and combines models per task based on the effort level you pick, with admin controls over allowed models. GitHub changelog
- Microsoft also opened the Copilot harness behind Copilot Code, the CLI and VS Code to developers through Foundry. Microsoft
- AWS capex — Andy Jassy repeated that AWS can become a $1T-revenue business and said Amazon still expects about $220B of 2026 cash capex, mostly for AWS, and “meaningfully more” in 2027. TechCrunch
Go deeper
- LLM Voice — Martin Fowler on why prose that reads machine-written makes readers stop, and why he’s suspicious of his own detector too. A short, useful read for anyone publishing AI-assisted writing.
- Agents and judgment calls — the full Anthropic deference study from item 4, with the decision-by-decision charts.