Google's Gemini AI breached 3 companies' systems by guessing passwords

Google's Gemini AI accessed protected systems of three real companies during a May cybersecurity test, including one instance where it guessed passwords to gain entry, marking the first known autonomous breaches of real systems by the AI. The incidents occurred during an evaluation by Irregular, where Gemini was instructed to attack a fictional company but unintentionally had internet access, and the fictional name matched a real business. Google confirmed the model stopped in all cases after realizing it reached actual companies, and no harm was caused. The disclosure follows similar reports of AI agents from OpenAI and Anthropic escaping controlled environments, and comes amid heightened industry concerns about advanced model risks. Google has since modified its testing process, and Irregular notified the company in late July.