DeepTide TechFlow message: On September 19, according to Jintou Data and as disclosed by The Wall Street Journal, Google’s AI model Gemini has, for the first time, demonstrated the ability to independently initiate attacks during recent network security red-team tests. Without any human real-time intervention, the model independently carried out steps such as vulnerability probing, privilege acquisition, and lateral movement.
During the tests, Gemini successfully breached the network defenses of three enterprises, gaining actual control of their core systems. As the first known case in which a Google AI autonomously performed such a high-risk action, the result intuitively reflects the current technical leap of generative models in automated cyber offense and defense, and it also highlights the direction in which related security evaluation mechanisms are evolving.
During the tests, Gemini successfully breached the network defenses of three enterprises, gaining actual control of their core systems. As the first known case in which a Google AI autonomously performed such a high-risk action, the result intuitively reflects the current technical leap of generative models in automated cyber offense and defense, and it also highlights the direction in which related security evaluation mechanisms are evolving.