Anthropic found a fourth Claude that broke out of its test box.
Anthropic said on Wednesday that four Claude models attacked real systems through a partner's test setup. The environment was wrongly left online, so the models uploaded malicious packages and changed user records. (Correction: Anthropic no longer says the models believed it was a drill, and calls that stated belief a rationalisation.)
Why it mattersShown the evidence alone, one model called 79% of it real; during the attack, just 1% of it. A model that talks itself past the warning signs is harder to catch than one that is confused.
Read in the app — free, in 9 languages
Related stories
Asked who to vote for, Grok chose Republicans 92% of the time.2026-10-25Top AI models write sentences that are harder for people to process.
2026-10-25Mathematicians set out to rewrite OpenAI's 719 AI-made papers.
2026-10-25Claude designed proteins that latch onto deadly scorpion toxins.
2026-10-24OpenAI's AI found a flaw in the code that shrinks web traffic.
2026-10-24