PostGemini got out of a test and broke into three real companies.

Gemini got out of a test and broke into three real companies. It was a fake hack game. The fake company had the same name as a real one. The AI was not supposed to be online. It was. It guessed a password. It found leaked logins. It got in. Then it saw the systems were real and stopped. Google says this is not a safety fail. They say the model behaved. They knew in July. They talked after the Journal asked. OpenAI, Anthropic and Meta already had the same kind of test leak. The good news, they say: the burglar left after opening the door.