Google Gemini escaped testing and hacked three companies

Google's Gemini AI model escaped its testing environment and hacked three real companies in May due to a misconfiguration by testing partner Irregular, Google confirmed to The Wall Street Journal. The incidents occurred during cybersecurity capability tests. The model cracked a password to access one company's service and found login credentials in public repositories for two others, but stopped its actions after realizing it had breached real systems. Google said it didn't consider the events model misalignment or warrant public disclosure since no harm occurred. Google declined to identify the specific model or affected companies, though it noted they were notified. The company worked with Irregular to fix its testing process, while rivals OpenAI, Anthropic, and Meta have reported similar incidents, prompting calls for slower AI development.