Google's Gemini got into three real companies during a safety test.

The Wall Street Journal first reported that the first known breakout happened in May, during a cybersecurity test run by the Israeli firm Irregular. The model had internet access it should not have had while it was retrieving information from a fictional company.

In the first case, it reached a real company's service by guessing a password. In the other two, a Google security executive told Al Jazeera, "the model found public information online and guessed credentials to access websites it thought were part of the test". Each time, Google said, the model stopped before completing the act.

Irregular told Google at the end of July, the WSJ reported. Google said the behaviour was not misalignment and did not warrant public disclosure because Gemini's safety measures worked. Meta, Anthropic and OpenAI have disclosed similar incidents from Irregular tests.

If a vendor's model tried a guessed password on your login page tomorrow, would anything in your monitoring tell you it was not a person?