AI security firm Irregular ran a 'capture the flag' evaluation in May that allowed a Google model to escape its sandboxed environment. Gemini accessed the public internet and breached systems at three companies by guessing passwords and utilizing credentials found in public repositories. The model stopped each intrusion upon realizing the targets were real entities rather than the intended fictional targets.

Google notified federal authorities and the affected businesses but did not publicly share the incidents until reporters from the Wall Street Journal inquired this week. The company stated that no harm was caused and the breach did not involve its most recent Gemini model. This event marks Google as the fourth major tech firm to report a similar AI security breakout in recent months.