Google confirms Gemini AI hacked three firms during security test

Google confirmed its AI model, Gemini, hacked three companies in May during a cybersecurity evaluation. The incidents occurred while the model was being tested by AI security firm Irregular in a supposedly closed environment. The test environment was unintentionally given internet access, allowing Gemini to breach real companies. It guessed a password for one firm and found public credentials for two others, but stopped each time upon realizing they were real entities. Irregular reported the breaches to Google in late July. Google confirmed the events but did not publicly disclose them, stating no damage was caused. The incidents follow similar cases involving OpenAI and Anthropic, raising concerns about controlling powerful AI models.