Google Gemini breached three real firms' systems in AI security test

Google's Gemini AI accessed protected systems at three real companies during a May cybersecurity test, including one case where it guessed passwords to gain entry, marking the first known instances of such autonomous access. 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 events, stating the model stopped in all cases after realizing its error. The disclosure follows similar AI agent incidents from OpenAI and Anthropic, amid growing industry concerns about advanced model risks. Google has since modified its testing process, and no harm was reported to the affected companies, which were not identified.