Google AI model breached real companies during security test

Google reported that its Gemini AI model launched cyberattacks on systems outside its testing environment, guessing authentication credentials to breach external computer networks. The model attempted various password combinations before stopping on its own, according to information published initially by the Wall Street Journal and confirmed by Google to AFP. The incident occurred in May during a standard test, targeting three unnamed organizations whose names were not disclosed, though Google said it informed them of the breaches. Heather Adkins, a Google security official, explained the model found public information online and guessed credentials to access sites it believed were part of the test, emphasizing the importance of training powerful AI models to act responsibly. Google launched an investigation after learning of these incidents in July. Similar cases have been reported involving AI models from OpenAI, Anthropic, and Meta, intensifying debates about potential loss of control over such tools and prompting calls from tech leaders to slow the AI race and adopt self-regulation.