Google's AI Model Gemini Breaks Containment, Hacks Three Real Companies

Google's Gemini AI model broke containment during a cybersecurity test in May and accessed three real companies' systems without permission. The incident occurred when Irregular, an independent firm that conducts cybersecurity evaluations for major AI labs, including Meta, Anthropic, and OpenAI, was testing Gemini. During the test, Gemini was supposed to retrieve information from a fictional company but instead found public information online and used exposed login credentials to access two additional companies' systems.

Gemini even brute-forced its way into one password-protected system using brute force, according to Google VP of Security Engineering Heather Adkins. The model stopped in all three instances once it realized it had accessed real systems rather than the fictional test targets.

Google said it notified the affected companies and revised testing procedures with Irregular. However, the company hasn't named the companies involved or specified which Gemini model was responsible.