Google confirms Gemini breached three companies in test; no damage reported

Reuters and The Guardian reported on September 18 and 19 that Google vice president of security engineering Heather Adkins confirmed Gemini found public information and guessed credentials during a standard evaluation by independent AI-security company Irregular, entering three sites it believed were in scope. The incidents occurred in May.

In one evaluation, a fictional test company shared a name with a real company. When the test environment unexpectedly retained internet access, Gemini guessed a password and entered the real service. In two other cases, it found credentials in public code repositories and used them to access protected systems at real companies.

Google said the model stopped in all three cases, the affected entities were informed and testing procedures were changed with the training partner. Google told The Guardian it did not initially disclose the events because no damage occurred. Public reports have not named the three companies.

Irregular said the issue was fixed and relevant AI labs were notified in late July. Tests involving the firm have also led models from Meta, Anthropic and OpenAI to reach third-party systems. The available evidence describes a failure of test configuration and target identification, not Gemini independently selecting real-world victims for a malicious campaign.