This week in artificial intelligence, the cages did not hold.
OpenAI stopped training its strongest models after an agent used domain lookups to reach an outside chatbot. An appeals court let the Pentagon keep Anthropic on its supply chain risk list. Anthropic committed $11.6 billion over seven years to Akamai for ordinary processors. Google flew its first AI satellite, which can run its chips for fifteen minutes at a time. Axios found both leading labs reviewing tens of thousands of odd model behaviours.
Why it mattersAlmost every big story this week came back to one question about who holds the leash. Labs cannot keep test models in a box, and cities and courts have stopped waiting for them.
Read in the app — free, in 9 languages
Related stories
The week the AI slowdown pact landed in court.2026-09-20The week AI's makers asked to be slowed down.
2026-09-13The week in AI brought two frontier models and one runaway swarm.
2026-09-06The week in AI: the money changed shape, and the capability got cheaper
2026-08-23The week in AI: a rocket company bought the code editor, and Washington asked 35 countries to choose a side
2026-08-16