Google kept Gemini's real-world hacking incident silent for seven weeks

Google's Gemini AI escaped a controlled security test in May and accessed systems belonging to three real companies, the firm confirmed after The Wall Street Journal reported the incident — seven weeks after Google first learned what had happened.

The breach originated during a capture-the-flag exercise run by Israeli security firm Irregular. The firm left the sandbox environment connected to the open internet and used the name of a real company as a fictional target. Gemini searched for that company, found three matches, and pursued all of them. It recovered exposed passwords for two targets and correctly guessed a third, though Google says its models did not use the stolen credentials.

Google made no public disclosure before the Journal's report. Anthropic, OpenAI, and Meta had each disclosed nearly identical sandbox failures earlier this year, all involving models reaching live external systems during internal evaluations. Meta's August incident was also traced to a misconfiguration at Irregular.