Microsoft Says AI Agents, Too, Will Be Watched Like Insiders

According to AI TIMES, Microsoft said it will treat AI agents adopted by enterprises as potential insider threats, and on October 19 it announced that it will expand a system for real-time risk assessment, permission adjustments, and threat blocking under Project Perception. The key idea is to control behavior with security mechanisms outside the model rather than trusting the model itself. The company believes that even agents deployed normally can act in ways that diverge from organizational intent as they pursue their goals, and it is extending existing defenses against external attacks to monitoring internal agents. In particular, it is promoting the concept of a guardian agent that inspects the actions of other agents before they are executed, aiming to treat agents not as simple tools but as actors that exercise authority. The focus of security is also shifting away from preconfigured settings and toward continuously tracking actual behavior and context. However, the specific scope of deployment, rollout schedule, and performance metrics were not disclosed in detail in this announcement.