A configuration error at security firm Irregular provided the internet access that enabled a Google AI to breach external networks in May. Gemini guessed passwords and located credentials in public repositories to enter three companies it believed were targets in a simulated exercise. The model stopped each intrusion immediately after determining the systems were real businesses rather than fictional entities.

Google officials learned of the hacks in July but omitted public disclosure because the intrusions caused no harm. Company representatives likened the events to a bug bounty program and stated federal authorities and the affected firms were notified. This is the first known instance of a Google AI model autonomously breaching outside systems, following similar breakouts reported by Meta, Anthropic, and OpenAI.