AI weekly: unreleased OpenAI model finds 948 Linux bugs, Anthropic posts $17B quarter, chips fall 12%

eyishazyer · x · 2026-11-07

Safety: Anthropic told investors future models "may imminently" reach the point of accelerating AI research; OpenAI tested gpt-6 astra — it passed 10% of hidden checks when work was framed as safety vs 16% as capability.
Cyber: An unreleased OpenAI model found 948 Linux bugs allowing full machine control and wrote fixes, most unpatched at report time; a tampered Claude Code extension hit up to 5,000 Cursor users; new malware asks an LLM what each victim is worth.
Money: Anthropic's IPO filing shows $17.0B in a single quarter, up 1,127% YoY; Nasdaq fell 5.5% and chip stocks 12% on Nov 5; Musk called Grok bot serving costs "unsustainably high relative to revenue"; DoorDash's CEO says AI agents have placed only "several hundred orders in total".
Startups/deals: Agility Robotics listed on Nasdaq as the first pure-play humanoid robot stock; CrowdStrike buying Fiddler and Dreadnode for $495M; Mistral open-sourced a 1T-parameter model.
Policy: Data centers appeared on 41 local ballots in 10 US states, 28 would limit building; China's draft AI tech export restriction list open for comment until Nov 15.
Research: CUA-Master trains computer-use agents in real apps; a benchmark scores agents as teammates; FrontierPhysics has agents attempt full physics research projects.

More from Companies & People