Google says Gemini models reached real company systems during May 2026 test

Google has confirmed that Gemini models accessed three real companies' systems during a May 2026 test, following a Wall Street Journal report.

The exercise was run by cybersecurity firm Irregular as a closed-environment "capture the flag" test of the models' cybersecurity abilities. The AI was told to pull information from a fake company that shared a name with a real one. A misconfiguration let Gemini reach the internet, where it went after real infrastructure instead of the fakes.

In one case, Gemini guessed passwords until it got into a company's online services. In the other two, it searched public software repositories and found login credentials that companies had accidentally left exposed.

Google's models reportedly stopped in all three runs once they realized they had hit real company servers. Irregular then changed its configuration to block internet access. The firm did not initially treat the event as worth further investigation and only told Google in July, after other AI hacking incidents made news. Google has since notified the affected companies.

Source: Ars Technica