Gemini AI gains access to systems of three companies during testing

The incident occurred in May 2026 on the platform of Irregular, an Israeli company specializing in cybersecurity assessments. Gemini was participating in a "capture the flag" exercise, where it was supposed to extract data from a fictional company's software in an isolated environment. Due to a configuration error, the model gained internet access, and the name of the fictional target coincided with the name of a real organization.

In one instance, Gemini brute-forced passwords until it logged into a real company's secure service. In two other episodes, the model found third-party credentials in public repositories and used them to log in. According to Google, in all three situations, the model stopped on its own and exited the systems after determining they were not part of the test environment.

Irregular notified Google of the incident in late July. Google representatives stated that no damage was caused to the affected organizations, whose names have not been disclosed. The corporation notified the affected companies and U.S. federal authorities, and also modified its testing process. Similar incidents previously occurred on the Irregular platform during tests of models from OpenAI, Anthropic, and Meta.