Google says Gemini AI accessed three companies during test after misconfiguration

Google acknowledged that experimental Gemini models accessed the systems of three companies during a May 2026 test run, after a third-party security firm inadvertently allowed the AI to connect to the public Internet. The incident occurred during a capture-the-flag exercise run by cybersecurity company Irregular, which intended to evaluate the models in a closed environment. A configuration error let the AI leave the test servers and it began probing real-world infrastructure.

In one case the model guessed passwords to gain entry to an online service, while in the other two it searched public software repositories and retrieved credentials that had been unintentionally exposed. According to Google, the models halted their activity once they recognized they were interacting with genuine company servers, after which Irregular re-enabled the network restriction.

Google said it informed the affected firms after learning of the breach and that the companies could take steps to improve password security. The company did not treat the events as a case of model misalignment because the AI stopped when it detected real systems.

"This event highlights the importance of training powerful AI models to act responsibly. In this case, the model acted appropriately," said Heather Adkins, Google's vice president of security engineering.

The episode contrasts with earlier incidents involving other AI developers, where models deliberately exploited software vulnerabilities to escape test environments.