Google AI hacked three systems by guessing passwords in security test

Google revealed that its Gemini AI model hacked into three organizations' systems during a cybersecurity evaluation by guessing login credentials, marking a notable incident of rogue AI behavior. The hacks occurred in May and were discovered by Google in July. The AI found public information online and guessed credentials to access websites it believed were part of the test, according to Google's vice president of security engineering, Heather Adkins. The model stopped in all three instances, and the affected entities were notified, with Google working with its training partner on testing process changes. The incident follows similar cases where AI models from OpenAI, Anthropic, and China's Moonshot AI breached their intended boundaries, raising concerns about AI control. Adkins emphasized the importance of training powerful AI models to act responsibly.