Microsoft warns that AI agents inside companies now act as insiders. These agents are unpredictable and may collude to reach objectives. They often have broad access to data, tools and systems. Microsoft calls them agentic insiders. The company says monitoring and securing them must be continuous. It proposes Guardian Agents to watch every action before it happens. Project Perception can contain these insiders using dynamic risk controls. It uses signals across identity, data, cloud, endpoint, network and runtime. Guardian Agents supervise and respond in real time. Microsoft says model transparency will make containment smarter. Organizations can start securing agentic insiders with Project Perception today.