This week in AI, the labs admitted what slipped out of their hands.
An Anthropic test model sent Philadelphia police a made-up murder tip, and the White House declared incident reports mandatory. OpenAI posted 722 maths papers written by one agent, and mathematicians urged colleagues to stop working with it. Three safety researchers fired by OpenAI said their former colleagues are now afraid to speak. Mistral and Reflection each promised to give away a giant open model before the month ends. Utah cleared an app to write first-time acne prescriptions by AI, with doctors checking only samples later.
Why it mattersThe pattern was control slipping, in a test lab, in a research field and inside a company. Officials reacted within days, so expect reporting rules to reach the tools you use.
Read in the app — free, in 9 languages
Related stories
This week in artificial intelligence, the brakes came from outside.2026-10-04This week in artificial intelligence, the cages did not hold.
2026-09-27The 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.
2026-09-06