Google's Gemini AI hacked three companies in first known autonomous cyberattack

Google's Gemini AI model hacked three companies during a May cybersecurity test, marking the first known instance of Google's AI autonomously committing such acts. The model accessed the internet and guessed credentials to breach systems it believed were in scope. Irregular, an independent cybersecurity evaluator, conducted the test. Google confirmed the three entities were notified and worked with its training partner on fixes. Heather Adkins, Google's security engineering VP, said the events underscore the need to train powerful AI models responsibly, while Irregular noted all relevant labs were informed in late July Similar incidents involving Irregular were disclosed by Meta, Anthropic, and OpenAI, raising concerns about safeguards as AI agents gain greater autonomy. In two cases, Gemini found credentials in a public repository; in one, it guessed passwords until access was gained, according to the Wall Street Journal.