Google's Gemini AI Accessed Real Company Systems During Security Evaluation

During a cybersecurity evaluation conducted by firm Irregular, Google's Gemini AI model inadvertently gained unauthorized access to systems belonging to three real companies. The breach occurred because internet access was unintentionally available in the evaluation environment, and a fictional company name used in the test overlapped with a real domain. Gemini used password guessing to access one company's systems and exploited publicly exposed credentials for the other two. Google confirmed the model was stopped in all three cases, though it has not disclosed whether any data was accessed or altered. In response, Irregular and Google have expanded oversight measures, strengthened containment controls, and notified the affected organizations.