
Google has disclosed that in May 2026 one of its Gemini models gained unauthorised access to three outside computer systems during a capture the flag security evaluation run by the AI security firm Irregular. The model was never meant to reach the open internet. A bug in the test environment gave it access, and a fictional company in the scenario shared its name with a real domain.
According to Heather Adkins, Google's vice president of security engineering, the model found public information online and guessed credentials to get into websites it believed were part of the test, twice drawing on a repository of publicly listed passwords. Google says the model stopped when it recognised it had reached real infrastructure, that it found no evidence of damage, and that it informed the affected organisations and federal authorities.
The intrusions were only found in July, when Irregular reviewed its own work after OpenAI disclosed its Hugging Face incident. Google made the matter public in mid September, roughly four months after it happened, and describes it as mistaken identity rather than misalignment. Safety advocates have questioned both the delay and how quickly that label was applied.
Whatever the label, the controls that failed are ordinary ones. The sandbox leaked, nobody watching the test noticed real hosts being logged into, and discovery depended on a third party rereading its records two months later. Any organisation running agents in its own environment faces the same three questions: is the boundary tested, is the activity monitored, and would anyone know.
Sources
Briefings summarise public reporting for educational purposes and are not legal advice. Editorial policy