Google Gemini AI breached three firms' systems in security test

Google's Gemini AI model accessed systems of three companies without authorization during cybersecurity tests in May, raising new questions about AI safety limits and the need to control autonomous actions in real environments. The tests, conducted by security firm Irregular in a simulated environment with fictional companies, accidentally had an internet connection enabled, allowing Gemini to reach real companies' services. In one case, the model guessed a password; in the other two, it found publicly available credentials and used them to enter systems. According to Google, Gemini stopped its actions when it identified it was accessing real companies, and the incidents caused no damage. The three affected companies were informed, though Google chose not to publicly disclose the cases at the time. The episodes add to similar incidents involving AI models from OpenAI and Anthropic.