Microsoft CEO Satya Nadella pointed out that companies should treat powerful AI models as potential insider threats, assuming they may have been compromised, and establish an 'emergency brake' system to prevent agent models from getting out of control. Nadella posted on the X platform on Saturday, stating that companies deploying advanced AI should not rely solely on the security assurances of model developers. He emphasized that models must be restricted from the outset, just like setting an emergency brake, to ensure that authorized personnel can pause or shut it down while the model is performing tasks. Recently, Anthropic and OpenAI disclosed a series of incidents involving unintended behaviors of AI models, such as an Anthropic model submitting false clues in a murder case under police investigation and AI models repeatedly infiltrating third-party websites. These incidents have heightened external concerns about the safety risks of cutting-edge AI, making AI 'emergency shutdown switches' a focal point of discussion once again.