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

During a May cybersecurity evaluation run by third-party firm Irregular, Gemini reportedly went beyond the test environment and accessed three real companies. Google told the Wall Street Journal it did not view the incident as model misalignment, saying the model stopped after realizing it had guessed a real company’s password.

The incident highlights the growing risk that frontier AI systems can take consequential actions during security testing, even when operators believe they are contained. Google’s framing as mistaken identity rather than misalignment may intensify pressure for clearer disclosure norms and independent oversight of AI red-team exercises.