← The Weekly AI Recap

This week in AI: 948 Linux security holes in one AI run

November 8, 2026 · 4 min read

Good morning,

OpenAI showed how fast AI can now break and fix software, voters and investors turned on the AI building boom, and an outside auditor says AI is speeding up Anthropic's own research.

OpenAI's security model found 948 ways to take over Linux

Tux the Linux penguin sits atop a long sunlit stone wall dotted with small lime pins marking loose bricks, while a lens bearing the OpenAI knot inspects it.

OpenAI launched GPT-6 Cyber on Thursday, a model built to find security holes and prove they can be used.

The headline result came from a single 10-day run against the Linux kernel, the core of most servers:

  • Found: 948 distinct bugs that let a low-level user take full control of a machine.
  • Cost: about $1.2 million in API credits, roughly $1,300 per working attack.
  • Firefox: 81 previously unknown high-severity flaws. Mozilla has already shipped fixes for many.

OpenAI rates the model "Critical" for cybersecurity, its highest risk level. Only selected security teams in its Daybreak Red program can use it.

These are OpenAI's own numbers, and the run still needed people. In an early test, 20 of 200 claimed attacks had worked through a different bug than the one assigned. When the scan ended, only 61 of the holes were patched upstream, with 43 more fixes staged.

OpenAI's advice is blunt: a well-resourced attacker with any account on an out-of-date machine should be treated as having full control. Turn on automatic updates for servers and laptops, and shorten your patch windows.

AI had a bad week at the ballot box and on Wall Street

A crane holds a data-center block motionless above an empty building plot while a lime ballot drops into a ballot box in the foreground.

Tuesday's US midterms became a vote on data centers. Voters backed restrictions in 40 of 41 local ballot measures, including bans in Ohio and moratoriums in Maine.

Markets took it badly. On Thursday the Nasdaq fell 5.5%, its worst day since April 2025. Nvidia lost more than 9%, about $450 billion in value.

  • The bet: Democrats won in Texas, Ohio and Georgia, the states with the biggest data-center plans, and investors expect slower construction.
  • Other factors: a weak Qualcomm forecast and rising interest rates.
  • Private money: OpenAI closed an $11 billion first tranche at a $1.25 trillion valuation, according to reports. It first aimed for $30 billion at $1.4 trillion.

If you're raising money this winter, expect investors to push harder on valuation.

An outside auditor says AI is speeding up Anthropic's research

A large gear bearing the Claude mark drives a smaller gear spinning faster, while a lime magnifying lens inspects the machine and a cream curtain hides part of it.

METR, a nonprofit that tests AI models, spent 11 weeks examining Anthropic's internal data and interviewing its staff. Its report is a rare independent attempt to measure whether AI is speeding up AI development.

The main findings:

  • Pace: Anthropic's models are improving about twice as fast as in 2024-2025, though METR says the true figure could be 1.3x or 3x.
  • Cause: mostly Anthropic's heavy use of AI in its own research. METR is confident faster growth in training compute did not cause it.
  • How: AI mostly assists researchers today. Fully autonomous work plays a smaller but growing role.

METR measured well-defined tasks, so this doesn't show AI improving itself without people. Anthropic also redacted most of the evidence METR found most useful.

Expect new Claude models to keep arriving faster than they did last year.

Quick hits 🗞️

  • Claude Code could send your email address to websites. When a site asked for contact details, the agent sometimes put your account email in its requests. Anthropic says a full fix lands early this week; until then, update to 2.1.317 or later. Source

  • The best open model is now free to download. Chinese AI lab Moonshot released the Kimi K3.1 Pro weights, though its 1.56 terabytes of files need a data center. Thinking Machines promises its smaller Inkling 2 Small next week. Source

  • Claude can now read your bank accounts. Personal Finance is rolling out in beta to US Pro, Max and Ultra users. Access is read-only through Plaid, and chats in the finance hub are not used for training. Source

  • ChatGPT now makes GIFs. Images 2.6 generates images 50% faster and can animate them. It is rolling out on all ChatGPT plans and in Codex, with two new image models in the API. Source

  • Any app can now plug into Siri AI. Apple opened a developer preview in the iOS 27.3 betas, so apps can expose actions like orders and bookings. App Store submissions open in spring, and Siri AI is still unavailable in the EU. Source

Model leaderboard 🏆

Via Artificial Analysis, which tests models independently.

  • Smartest:
    • Gemini 4 Carbon (60), cyber defenders only for now
    • Claude Opus 5.5 (59)
    • GPT-6 Astra (57)
  • Best value: GPT-6.1 Terra (51) at about $2 per million tokens, still the cheapest top-10 model per task.
  • Best open-weights: Moonshot's Kimi K3.1 Pro (46), free to download.
  • Moved this week: Kimi K3.1 Pro took the open-weights spot once its weights went public.

See you next week!

If this helped you catch up quickly with AI, forward it to someone who would find it useful. They can subscribe here.