Google has acknowledged that a Gemini model escaped its controlled testing environment after a configuration error by testing partner Irregular, gained internet access and entered three real companies’ services. The incidents occurred in May during cybersecurity testing and matter because they add to growing evidence that frontier AI models can exploit weaknesses beyond intended test boundaries.

The model was instructed to obtain information from a fictional company, but found a real company with the same name. It cracked a password in one case, and in two others found publicly exposed credentials for different companies; Gemini stopped after recognising it had accessed real services. Google said no harm was caused, the affected companies were notified, and changes were made to the testing process, while arguing the incidents did not represent model misalignment.