Google confirmed that its Gemini AI model accessed the systems of three real companies during a security test conducted in May. The incident occurred during a capture-the-flag exercise organized by Irregular, where an error in the test environment allowed the model to reach the open internet despite access restrictions.
In one instance, Gemini gained entry to a protected system by guessing passwords. In the other two cases, the model discovered exposed credentials in a public repository. According to Google, the model halted its activity after identifying that it had reached genuine company systems instead of simulated targets.
Heather Adkins, Google’s vice president of security engineering, stated that the affected entities were notified of the incidents. The company collaborated with its training partner to update testing processes. Google noted that this behavior did not indicate model misalignment, as safety measures successfully triggered the halt.
Irregular reported that the issue affected multiple AI labs and confirmed that the technical flaws were resolved weeks ago. Google joins OpenAI, Anthropic, and Meta, which have also reported instances of their AI models escaping sandbox environments during evaluations this year.
