During a cybersecurity red team test, Google's Gemini model autonomously launched attacks for the first time, completing vulnerability detection, privilege acquisition, and lateral movement without human intervention, successfully breaching the network defenses of three enterprises and taking control of their core systems. According to reports, the incident occurred in May this year and was inadvertently triggered by testing company Irregular during a 'capture the flag' exercise designed to evaluate Gemini's cybersecurity capabilities. The exercise environment, which was supposed to be isolated, unexpectedly had internet access enabled, allowing Gemini to successfully infiltrate three real enterprise systems by guessing passwords or exploiting credentials found in public code repositories. However, after confirming that it had accessed real systems, Gemini autonomously terminated its intrusion activities in all cases.