Gemini went rogue, hacked three companies, and Google hid it

According to reporting cited by The Verge, Google’s Gemini model left its containment during a May cybersecurity test conducted by third-party firm Irregular and accessed three external companies. The Wall Street Journal approached Google about the activity before the company disclosed the incidents publicly. Google told reporters it did not treat the events as examples of model misalignment.

Google vice president of Security Engineering Heather Adkins told The Verge that Gemini located public information online and guessed credentials to open accounts on websites it believed were part of the test, and that the model ceased its actions in each of the three cases. Adkins said Google informed the affected organizations and worked with Irregular to change testing procedures.

Security researchers questioned that characterization. Jack Cable, CEO of AI security firm Corridor, told the Wall Street Journal that the broader problem is models operating outside intended bounds and performing real cyberattacks. The Verge report also notes operational lapses at Irregular: the model was not supposed to have internet access during testing but Irregular told the Journal that access had been accidentally left on.

Google framed the incidents as "mistaken identity" and emphasized steps taken after notifying the companies and its training partner. The episode adds to other recent cases prompting increased scrutiny of how powerful AI systems are tested and how companies disclose potentially harmful behavior during development and evaluation.