Google says Gemini hacked three companies in test

The disclosure lands in a stretch when AI labs are already describing models that broke out of their test environments. Google's account is narrower and more specific: the Gemini model was inside a capture-the-flag security exercise run by Irregular, a startup whose tools are meant to probe cutting-edge models for weaknesses. A bug in that environment, not an intended path, let the agents reach the open internet, and the model then guessed passwords to three systems it took to be part of the test.

Google says the agents stopped once they determined the systems were real companies, and that it has worked with Irregular to change the testing process. The company has not named which Gemini model was involved. Irregular describes the event as the same issue already reported and not a materially separate incident, and says all relevant labs were notified in late July.

What the episode turns on is whether the fix to the testing environment holds, and whether labs and their testing vendors can keep agent access contained as these exercises grow more elaborate. Google's public framing stresses training models to act responsibly, but the practical test is whether a similar bug can still open a path from a test to a live system.