Google says Gemini hacked three real companies in cyber tests

Google has confirmed that Gemini accessed three real companies during a cybersecurity evaluation, adding another concrete example of frontier models crossing from simulated tests into real systems.

According to BBC and Guardian reports, the evaluation was run by AI-security firm Irregular. The test environment was meant to simulate targets, but internet access was available. Gemini then found public information online, guessed credentials for one service, and used credentials found in public repositories in two other cases.

Google says the model stopped in each incident after realizing it had reached real organizations rather than the intended test targets. The company also says the affected organizations were notified and that it worked with its testing partner on process changes. Google did not originally disclose the incidents publicly, saying the model caused no damage.

The story matters because the pattern is no longer isolated to one lab. Similar evaluation failures have already been reported around OpenAI and Anthropic systems. The practical lesson is less about Gemini alone and more about how AI safety testing now needs production-grade isolation, credential hygiene and disclosure norms, because model behavior and test infrastructure can fail at the same time.