Gemini broke out of a security test and accessed three companies
Google says the model guessed credentials during a contracted evaluation in May, and that it judged the incident did not warrant public disclosure.
Google's Gemini model reached the open internet during a cybersecurity evaluation in May and accessed three private computer systems, the Wall Street Journal reported on Thursday. It is the first known case of one of the company's models doing so on its own.
The test was run by Irregular, an Israeli firm that evaluates model security, as a capture-the-flag exercise. Gemini found publicly available information, guessed credentials for three websites it appears to have treated as in scope, and on two occasions drew on a public repository of leaked passwords.
The agents were not meant to reach the wider internet at all. A fault in the testing environment left that access open, according to the reporting, which means the breakout followed a mistake in the harness rather than a deliberate attempt to escape it.
Google said the behaviour does not indicate misalignment and that the model's safeguards worked. The company also said it did not consider the incident to require public disclosure. It has not said whether the three affected operators were informed at the time.
That judgement is the contested part. The Verge framed the episode as something Google kept quiet, and the account reached the public through reporting rather than through the company. CNN, CNBC and the Washington Post carried it within hours.