Microsoft outlines continuous oversight model for enterprise AI agents
Company says agents it ships cannot be assumed safe, and presents Project Perception as a way to manage them.
Microsoft has set out its position that AI agents deployed inside companies are "Agentic Insiders" that require continuous oversight and control, and said its Project Perception system can be used to manage and control them, according to datanet.co.kr.
The company said agents it shipped itself cannot be assumed safe, because they may act unexpectedly or misuse excessive permissions. It also warned that reward hacking can lead a model to choose unintended methods to reach its goals.
Microsoft proposed "Guardian Agents" that review other agents' actions before execution and monitor them around the clock, applying controls when risk rises. It said agent protection must extend beyond the model to the harness, the tool-access components and the runtime, along with identity, data, endpoint, network and cloud layers.
Project Perception combines red, blue and green agent teams with a new cybersecurity stack to counter AI-driven attacks, and its protection structure covers three areas: sensors, signals and actuators; context; and harness, agents and models. The system provides traceability information so security staff can decide on responses and adjust policies.
Microsoft said automated agentic security and human judgment should complement each other, and named model transparency, including visibility into the reasoning process, as a challenge for protecting against agentic insiders. Hayete Gallot, Microsoft Security corporate vice president, said companies must grasp agent activity and risk in real time and continuously adjust action permissions.
Background
In July 2026, Microsoft announced Project Perception as an agentic security system built around red, blue and green agent teams to respond to AI-based attacks.
Quick answers
What does Microsoft mean by "Agentic Insiders"?
Microsoft uses the term for AI agents deployed inside companies, which it says need continuous oversight and control.
What are Guardian Agents?
Guardian Agents are agents Microsoft proposes to review other agents' actions before execution and monitor them around the clock, applying controls when risk rises.
What is Project Perception?
Project Perception is Microsoft's agentic security system, which the company says can be used to manage and control agentic insiders.