AGENTS are rapidly becoming part of how work gets done, and security must keep pace with their machine-speed capabilities. Microsoft describes Guardian Agents and Project Perception as a way to defend the agentic insider, with signals and a unified registry giving IT and security teams a consistent view of agents and risk. The approach emphasises surrounding non-deterministic intelligence with deterministic controls, continuous observation, and task-scoped authority.
The piece stresses that agents may be powerful yet can still be flawed, manipulated, or compromised, and that the trust architecture must separate intelligence supply from security authority.
Microsoft outlines four continuous responsibilities for securing agents: Know which agents exist and what they are doing; Contain what each agent can access and do before it acts; Protect the data, tools, and systems in the agent’s path; and Adapt protection as behaviour, context, and risk change.
They describe concrete mechanisms to implement this, including discovery of local agents, a durable agent identity separate from users, sandboxes such as the Microsoft Execution Container, and policy-driven controls that limit the agent’s capabilities. The article highlights Dynamic Permissions, Dynamic Agent Risk, and Dynamic Conditional Access as core features to grant least privilege and adjust access in real time, rather than treating risk changes as a binary decision.
The piece also discusses securing at scale, with Guardian Agents carrying out work at machine speed under human-defined policy, and the extension of model transparency to improve security evidence without letting the model override enforcement. The aim is to enable organisations to deploy agents with confidence, ensuring security observes, governs, and contains at the same speed as innovation. Published on 19 October 2026.