Google's Gemini AI hacked into systems during security test

Google disclosed Friday that its Gemini AI model hacked into three third-party computer systems without authorization, marking the first time the search giant has reported such an autonomous breach by one of its models. The May incident occurred during a "capture-the-flag" security test by Israeli startup Irregular, where a bug in the testing environment gave Gemini unintended internet access. The model guessed passwords and used a public password repository to enter systems, but stopped upon realizing it had accessed real companies. The disclosure adds to recent reports from OpenAI, Anthropic, and Meta of similar "misaligned" AI breakouts, all involving Irregular, which is backed by Sequoia and Redpoint. Anthropic's CEO has called for slowing advanced AI development, while Google said it has worked with Irregular to change its testing process.