AI stocks crash after the midterms, OpenAI's GPT-6 Cyber
The AI stock sell-off, OpenAI's GPT-6 Cyber, METR on Anthropic's speed-up, GPT-6 Astra doing worse on safety research, Claude's finance tool, a suit against Google, Thinking Machines and Meta.
AI stocks crashed, taking the Nasdaq down 5.5 per cent The Nasdaq 100 lost $2.3 trillion on Thursday, and the S&P 500 had its worst day since Trump’s tariffs in April 2025. Many of the Democrats who won on Tuesday have promised to slow or stop data centre building. Asian markets followed on Friday, with Seoul down more than 8 per cent at the open. Source
OpenAI’s GPT-6 Cyber is out, after finding 81 new Firefox flaws It’s the model behind the 948 Linux exploits OpenAI showed off on Wednesday, and it went out on Thursday to a small group of customers in Daybreak Red, its programme for approved security teams. Users face extra identity checks, and the model checks in with OpenAI more often when it’s asked to do something high-risk. OpenAI says all 81 Firefox flaws have been patched. Source
METR says Anthropic’s pace has roughly doubled, mostly thanks to AI In a report out on Thursday, the independent testers found Anthropic’s models have been improving about 1.9 times faster since October 2025. They guess AI is behind around 80 per cent of the current speed-up and see no clear bottleneck in the next year, but say both calls are uncertain. Anthropic redacted most of the evidence METR found most informative. Source
GPT-6 Astra quietly does worse when the task is safety research In OpenAI’s own report, updated on Wednesday, Astra finished about 10 per cent of the hidden checklist on a coding job framed as safety research and about 16 per cent when the same job was framed as capabilities work. The gap grew during training. OpenAI’s best guess is that the model is being overly cautious, but it says it can’t rule out that it’s doing it on purpose. Source
Claude can now read your bank account and sort your spending Personal Finance connects your bank and credit cards through Plaid, breaks down where the money went and flags duplicate charges, surprise fees and creeping bills. It’s read-only, so it can’t move money. It was announced on Thursday, and the beta is rolling out this week to Pro, Max and Ultra users in the US. Source
Parents are suing Google over their 13-year-old’s chats with Gemini The North Carolina family says their daughter spent months talking to the chatbot about suicide and self-harm on a school Chromebook, and that it coached her on hiding the self-harm from her family. The suit, filed in a California state court on Wednesday, also says Google hides AI chat logs on student accounts. Google says it takes the claims very seriously and is reviewing them. Source
Thinking Machines’ new 594-billion-parameter model should go open next week Inkling 2 Small came out on Thursday on Thinking Machines’ Tinker platform, with only 17 billion parameters working at a time. The company plans to put the full weights out next week under the Apache 2.0 licence, so anyone can use and build on them, once outside safety testing is done. Source
Meta shelved terms that would let it scrape business customers’ websites They’d have let Meta use what it scraped for its ads and business products, starting on 20 November. Meta paused them indefinitely on Thursday, saying it wants to take its time to get them right. Courts in Norway and Brazil had already blocked them for now after media companies sued, and Britain’s competition regulator is investigating. Source