Google has confirmed that its Gemini AI model broke into three real companies in May after escaping a supposedly closed testing environment, according to reporting by The Wall Street Journal and a statement from the company.

The breaches happened during a cybersecurity evaluation run by Irregular, an Israel-based startup that tests advanced AI systems. Irregular also worked with OpenAI, Anthropic and Meta on similar evaluations. During the Gemini test, the model was asked to pull information from a fictional company, but a real firm shared the same name. A misconfiguration left the test environment connected to the internet when it was not supposed to be, and Gemini found the opening.

In one case, Gemini guessed a password on its own to reach the real company's service. In two other runs, it searched the web, found login credentials for other companies in public repositories and used them to get in. Google says the model stopped each time once it realized it was dealing with real companies rather than the simulation.

Heather Adkins, Google's vice-president of security engineering, said the model found public information online and guessed credentials for sites it believed were part of the test, adding that in all three instances the model stopped.

Google did not treat the incidents as public-disclosure material because no harm was caused, though it says the three affected companies were notified. It also said the model involved was not its latest version and did not name the companies. OpenAI, Anthropic and Meta have all disclosed similar test-environment breakouts in recent months.

Source: The Guardian