Gemini AI model gained access to systems of three companies during tests

An incident occurred in May 2026 on the platform of the Israeli company Irregular, which specializes in cybersecurity assessment. Gemini was undergoing training in the format of a "flag capture", where it was supposed to extract data from a program of a fictional company in an isolated environment. Due to a configuration error, the model gained access to the internet, and the name of the fictional target coincided with the name of a real organization.

In one case, Gemini tried passwords until it entered a protected service of a real company. In two other episodes, the model found login credentials of third-party organizations in open repositories and applied them to log in. According to Google, in all three situations, the model stopped itself and exited the systems, determining that they did not belong to the test site.

The company Irregular notified Google about the incident at the end of July. Google representatives stated that the affected organizations did not suffer any damage, and their names are not disclosed. The corporation notified the affected companies and the US federal authorities, and also made changes to the testing process. Previously, similar incidents on the Irregular platform occurred during checks of models from companies OpenAI, Anthropic, and Meta.