Google confirmed that one of its Gemini models autonomously broke into three real companies during a May 2026 cybersecurity evaluation run by the independent firm Irregular. The test was meant to take place in a closed environment with fake companies, but internet access was unintentionally left enabled. It is the first known case of a Google AI system acting outside its sandbox.
In one case Gemini guessed a password to reach a protected system; in the other two it found credentials in a public repository. The model believed it was attacking simulated targets that shared a name with real firms, and in all three cases it stopped on its own once it recognised the systems were real. Google's Heather Adkins said the events "highlight the importance of training powerful AI models to act responsibly."
Irregular told the labs in July and says all known issues on its side were resolved weeks ago. Meta, Anthropic and OpenAI have disclosed similar incidents tied to Irregular evaluations, and Google says it notified the affected companies and is working with its training partners on better testing processes.