Anthropic invites outside reviewers in

Dario Amodei published a three-step plan for slowing the pace of AI capability gains, and committed Anthropic to the first step on its own: an external review team with desks in its offices, access badges and company laptops, plus permissions mostly comparable to its own internal risk assessors. He names METR as the kind of group he means and compares the arrangement to the supervisors that regulators embed inside banks. His worry is specific. He writes that within six to twelve months, an agent swarm like the one behind the OpenAI Hugging Face incident could take over the internet with a persistent botnet.

Parts of Silicon Valley met the recent run of insider warnings with skepticism at a Goldman Sachs conference in San Francisco. Grindr chief executive George Arison told the BBC the comments reflect an "anti-civilisational worldview at Anthropic", called them dangerous, and said he has instructed some engineers to stop using Anthropic's technology. Others suggested the warnings are timed to talk up products ahead of record-setting listings. Anthropic, valued at $965 billion in May, had not responded when the piece published.

Asked by Fortune editor-in-chief Alyson Shontell whether OpenAI would list, Altman said that given everything happening with safety, now would be an ill-advised moment to go public, and that the company is not rushing. Pressed on whether that excluded this year, he said it did, adding that OpenAI has a lot still to do. The company had been reported to have hired bankers and lawyers against a target of late 2026 before leaning toward 2027.

Specific Labs licensed real production code from companies, including a billing service that has to get sales tax right, and set eight frontier models against the actual tickets those engineers work on. Fable 5.1 in Claude Code led at 38.8 percent, ahead of GPT-6 Astra at 33.8 percent and Gemini 3.8 Flash at 31.2 percent, and the most common failure was simply missing a requirement. Two caveats belong with those numbers: the leaderboard rests on a ten-task sample with eight runs per model, and the benchmark is the vendor's own.

Writing on his own site, the Turing Award winner argues the recent run of agent misbehavior has a common root rather than being a string of unrelated bugs. Training rewards outcomes, so systems learn to pursue whatever raises the score, including editing the files that grade them, and capable models can tell when they are under evaluation and act differently. He reads the OpenAI Hugging Face forensics much as Amodei does, and argues the answer is pacing and a change in how models are trained rather than a patch per incident.

Sequoia Capital is leading a deal that would value Mecka AI at $500 million, up from $60 million three months ago. The company pays people to wear body sensors and record themselves doing ordinary physical tasks, then sells the footage to train humanoid robots — the same trade Scale AI and Surge made supplying text to language-model builders. Founded in 2024, it forecast an annual run rate of $100 million by the end of this year.