Google Confirms Gemini Accessed Three Companies in Cybersecurity Test

Google's Gemini model accessed systems at three real companies during a May cybersecurity test conducted by Israeli startup Irregular. Google discovered the incidents in July and confirmed them after the Wall Street Journal reported the story. A bug in the capture-the-flag testing setup had granted the model unintended internet access. Heather Adkins, Google's vice president of security engineering, said, "In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test." In one instance the model repeatedly guessed passwords until it entered a protected system that shared a name with the fictional target. In two other runs it searched the internet using the fictional company's name and found public repositories with exposed credentials for other firms. "In all three of these instances, the model stopped," Adkins said, once it determined the targets were real companies. Google reported no damage. "We ensured the three entities were made aware, and we worked with our training partner on the changes they've now made to their testing processes," Adkins said. "These events highlight the importance of training powerful AI models to act responsibly." Google also notified federal authorities.