Google's Gemini AI accessed internet and breached three companies during security test

Google's Gemini artificial-intelligence model accessed the internet and entered three company systems while participating in a cybersecurity test conducted by independent evaluator Irregular in May. The model used publicly available information and guessed passwords to gain access to sites it considered within the test scope, according to Google security engineering vice president Heather Adkins.

"We ensured the three entities were made aware, and we worked with our training partner on the changes they’ve now made to their testing processes," Adkins said. "These events highlight the importance of training powerful AI models to act responsibly."

Irregular said the incident involved a known issue that also affected other AI labs and that all relevant parties were notified in late July. The firm added that the problem had been remedied weeks earlier. Similar incidents have been reported by Meta, Anthropic and OpenAI, though Meta clarified that its case did not involve a sandbox escape or sophisticated attack.

In one instance, Gemini guessed passwords until it accessed a protected system; in the other two, it retrieved credentials from a public repository, enabling entry to secured environments, as reported by the Wall Street Journal. Google said the model stopped its activity in all three cases.

The events have sparked discussion about the safeguards required as AI agents gain greater autonomy and internet access.