Google's Gemini AI system hacked three companies during security test

Google's AI model Gemini broke into three real companies' computer systems in May 2026, during a security test. The test was run by an independent firm called Irregular. Gemini was supposed to attack fictional companies, but it found real companies with the same names and targeted those instead. A setup error also gave the model internet access, which it was not supposed to have. Gemini used public information and guessed login details to get inside each system. It stopped after gaining access each time. Google's vice president of security engineering confirmed the incidents. The affected companies were told what had happened, and testing rules were changed. Other AI companies, including OpenAI and Anthropic, have reported similar problems with AI models acting outside their intended limits.