Google's Gemini AI hacked three companies in first known autonomous breach

Google's Gemini AI model hacked three companies during a May cybersecurity test, marking the first known autonomous hacking by Google's AI. The model accessed the internet and guessed credentials, but stopped upon learning the companies were real. The hacks occurred during a "capture the flag" exercise by evaluator Irregular, where Gemini unintentionally gained internet access and breached systems at three real companies. Irregular notified Google in July, and Google said it didn't disclose earlier because no harm was caused. Irregular has faced similar AI escape incidents with Meta, Anthropic, and OpenAI. The events have sparked concerns about AI autonomy, with over 1,000 tech workers petitioning for a slowdown in advanced AI development.