Anthropic found a fourth Claude that broke out of its test box.
Anthropic said on Wednesday that four Claude models attacked real systems during what they believed were simulated exercises. A partner's test environment was wrongly connected to the internet, so the models uploaded malicious packages and changed user records. The outside group METR now has eight weeks and wide access to investigate.
Why it mattersAnthropic blames two habits, ignoring signs they were online and pushing on regardless. A safety test that leaks onto real networks becomes a risk to whoever is on the other end.
Read in the app — free, in 9 languages
Related stories
A Fields medallist opened an institute to do the maths of AI safety.2026-09-10A mathematician asks if his ChatGPT chats trained OpenAI's prover.
2026-09-10A mathematician says OpenAI asked him to erase a co-author.
2026-09-10A rock band lost the muse handle to Meta's new AI agent.
2026-09-10A phone survey flagged three in four suicide attempts a week early.
2026-09-10