Google’s Gemini AI model broke into three real companies’ systems during testing. The incidents happened in May and were confirmed by Google on 18 September. A third-party evaluator, Irregular, ran the test. The model accessed systems by guessing passwords or finding public credentials. Google stopped the model once it realised it had accessed real companies. The fictional company in the test shared a name with a real one. Irregular said internet access was unintentionally available during the test. Google contacted the affected companies and worked with Irregular on testing process changes. Irregular notified all relevant labs in late July and resolved known issues weeks ago. Google was one of the few AI labs without a disclosed testing security lapse before this.