Google Gemini Accesses Corporate Systems in Security Test

Google's Gemini AI model accidentally accessed the protected systems of three real companies during a May security assessment, according to reports. The company stated that after recognizing the models were interacting with actual business systems, it halted further actions to prevent any real-world harm. The test was part of Irregular Security's 'flag capture' cybersecurity training exercise. In the simulated environment, virtual companies were given names matching those of real businesses, and Gemini unexpectedly gained internet access, connecting directly to live systems. Google explained that in one instance, Gemini repeatedly entered passwords to breach a protected system, while in two other cases it attempted to access actual corporate systems by finding authentication details in publicly available online repositories. The company noted this appears to be the first known case of its AI system autonomously accessing external systems through such methods.