Google Gemini AI hacked 3 real firms during security test, then stopped itself

Google's Gemini AI model accessed three real websites during a cybersecurity evaluation in May, using public information and guessed credentials, marking the first known autonomous hack by the company's AI. Google confirmed the incident Friday, stating Gemini halted after realizing the firms were real. The test, conducted by independent firm Irregular, tasked Gemini with targeting a fictional company, but it instead breached real organizations. Google's Heather Adkins said affected entities were notified, and Irregular confirmed all labs were informed in late July, with issues since remedied. The incident has intensified debates on AI safeguards as agents gain autonomy, with industry leaders like OpenAI's Sam Altman and Google DeepMind's Demis Hassabis advocating for regulatory oversight amid warnings about advanced AI risks.