Google says its Gemini AI model gained unauthorized access to three external computer systems during a cybersecurity test, marking a notable incident involving an AI system operating beyond its intended environment.
According to Google, the incident occurred in May while Gemini was undergoing testing. The model either guessed login credentials or discovered publicly available credentials and used them to access three systems it mistakenly believed were part of the evaluation.
Google said Gemini stopped before taking further action and that the unauthorized access caused no known damage. The company does not classify the incident as AI “misalignment,” arguing that the model mistakenly identified real-world systems as part of its testing environment.
The issue was discovered in July after cybersecurity company Irregular reviewed the tests following reports of similar incidents involving other AI systems. Google then investigated, notified the affected organizations, and informed federal authorities.
The incident adds to growing scrutiny around increasingly autonomous AI agents. OpenAI and Anthropic have also disclosed cases in which their models demonstrated unexpected behavior during cybersecurity evaluations.
Google said the incident highlights the importance of stronger safeguards and responsible training as AI systems become increasingly capable of interacting with real-world computer infrastructure.