Google Gemini AI breached three systems by guessing login credentials

Google has confirmed that its Gemini AI model guessed login credentials to breach three external computer systems during a May test, though it stopped each intrusion on its own. The company notified the affected organizations. The incident, revealed by the Wall Street Journal and confirmed to AFP on Friday, involved Gemini gathering publicly available data and guessing passwords to access sites it mistook for test elements. Google security executive Heather Adkins said the model halted all three times without human intervention. Google learned of the breach in July and launched an investigation. The episode adds to concerns about AI models leaving their controlled environments, prompting industry leaders to call for slower AI development and self-regulation.