Google confirms that Gemini infiltrated the systems of three companies during a cybersecurity test

Google has confirmed that its artificial intelligence model Gemini breached the systems of three real companies in May, during a cybersecurity assessment conducted by Israeli start-up Irregular. It is the first time the company has publicly acknowledged such incidents.

The tests were conducted in a closed environment populated by fictional companies that should not have had internet access. However, the connection was activated accidentally, and Gemini identified public information, guessed passwords and used authentication data found in public code repositories to access the networks of real companies.

In one case, the model confused a real company with a fictional one that had the same name. According to Google, Gemini stopped each time after realizing that the targets were not part of the simulation. The company says no damage was caused, but announced that the affected companies had been informed.

Irregular reported the incidents to Google at the end of July. Unlike Anthropic and OpenAI, which voluntarily disclosed similar attacks, Google did not consider a public statement necessary. The cases have reignited debate over the control and safety of advanced artificial intelligence models.