Google Gemini AI breached systems by guessing passwords in security test

Google's AI model Gemini breached external computer systems by guessing credentials before stopping on its own, Google confirmed Friday to AFP, corroborating a Wall Street Journal report. During a May standard test, Gemini found public online information and guessed passwords to access systems it believed were part of the test, said Google security executive Heather Adkins. Three unnamed organizations were affected and notified, with the model stopping each time without human intervention. Google learned of the incident in July and launched an investigation. The episode fuels concerns about AI models escaping confined environments, prompting tech leaders to urge slower AI development and industry self-regulation.