AI agents are becoming active participants in the enterprise. With access to data, tools and systems, they can take actions and make decisions across critical processes. This new generation of “Agentic Insiders” requires a new approach to security, with continuous monitoring, dynamic risk assessment and controls that can respond in real time.
Microsoft’s agentic security system Perception applies this approach by combining deterministic security controls with non-deterministic AI systems and supervising agent activity across the enterprise.
The most important key points:
- AI agents introduce a new insider risk: Frontier AI models are optimized to accomplish objectives, but they do not inherently understand an organization’s norms or judgment. As agents gain broader permissions and access to enterprise systems, organizations need to treat their behavior as a security risk even when an agent has not been compromised by an external attacker.
- Security needs to extend across the entire agent stack: Protecting agents requires controls spanning identity, permissions, data, endpoints, networks, cloud infrastructure and runtime environments, with visibility into what agents are doing and the risks their actions create.
- Perception brings continuous supervision to AI agents: Microsoft Perception brings together signals across the security stack to discover and manage agents, dynamically assess the risk associated with their activity and enable Guardian Agents to continuously supervise and respond as risk changes in real time.
- Agentic security combines machine speed with human judgment: Perception can provide real-time supervision and traceability, while security practitioners investigate what happened, understand why risks emerged and adjust security policies accordingly. The goal is to enable defenders to operate at the speed of the AI systems they protect.
- Greater model transparency is the next step: Looking ahead, Microsoft argues that security systems will increasingly need to understand not only what an agent is doing, but why it is doing it and whether its behavior remains aligned with the intent of the task. Greater model transparency can make containment, accountability and traceability stronger.
The central principle is that organizations should not have to choose between AI innovation and safety. As enterprises deploy more autonomous agents, security and governance need to evolve alongside them.
You can find the original post by Hayete Gallot, President, Security, Microsoft, here.