Pro User
  • All of Microsoft
  • My Role
    • All in My Role
    • Pro User
    • Administrator
  • Focus
    • All in Focus
    • Microsoft Ignite 2026
    • Microsoft Copilot
    • All about AI
    • M365 Admin
    • What's Hot
    • M365 Release
Timespan
explore our new search
Microsoft Entra Agent ID separates agents from users
Microsoft Entra

Microsoft Entra Agent ID separates agents from users

Microsoft Entra Agent ID gives supported local AI agents separate identities so IT teams can assess risk and apply agent-specific controls.

Key insights

  • Microsoft Defender and Global Secure Access help discover local agents; Entra assigns discovered agents an Agent ID.
  • Dynamic Agent Risk combines runtime, exposure, and blast-radius signals from Entra, Defender, and Purview.
  • Dynamic risk controls can allow, constrain, require human approval, restrict sensitive access, or block an agent.
  • The session identifies Entra Agent ID for local agents and Dynamic Agent Risk as preview features; Conditional Access dynamic controls are in limited preview.
  • An agent can be blocked without blocking the human user whose device it runs on.

Microsoft Entra Agent ID gives supported local AI agents separate identities, letting organizations distinguish agent activity from the actions of the user who runs them. This matters to identity, security, and IT teams that need to discover agents, understand their access, and respond to changing risk. Microsoft Events demonstrates the discovery and policy workflow in a YouTube session.

What's new

The security model presented here treats an agent as a distinct identity rather than letting it act with the same identity and permissions as its user. That separation gives administrators a way to identify agent activity and apply controls to the agent without automatically disrupting the person using it. The session covers agents from Microsoft and other providers, including agents running locally on devices.

For local agents, Microsoft Defender provides visibility into activity on endpoint devices, while Entra Global Secure Access (GSA) provides visibility at the network layer. Together, the session says, these services can discover more than 60 local agents. Entra can assign an Agent ID and record the agent type, while Agent 365 provides a registry view for discovered agents.

The other major change is Dynamic Agent Risk. It brings together signals from Entra, Defender, and Microsoft Purview into a shared risk view. The risk combines runtime behavior, exposure to exploitation, and blast radius—the potential harm based on the agent’s access. The signal can change as new activity is detected.

Conditional Access dynamic controls use that changing risk to shape what an agent can do. Instead of only allowing or blocking access, controls can constrain access, restrict sensitive actions, or require human approval. This makes it possible to respond differently to medium and high risk rather than treating every risk increase as an immediate full block.

Who it is for and what you need

These capabilities are aimed at IT administrators and security operators responsible for agent inventory, identity, and access policy. They are relevant when employees use local agents as well as when organizations manage agents across different platforms. The central operational point is to distinguish the agent’s identity and risk from the human user’s.

The session describes Entra Agent ID for local agents and Dynamic Agent Risk as preview features. It identifies Entra Conditional Access dynamic controls as limited preview. Local-agent discovery depends on Microsoft Defender and Entra Global Secure Access, which the session presents as part of the discovery workflow. The announcement slide specifically names supported local agents such as GitHub Copilot CLI and Claude Code.

The session also describes a wider discovery view that includes third-party agents, but the named supported local-agent examples are the clearest starting point for assessing the feature. Administrators should consider which agents are present, what resources they can reach, and how policy should respond to different risk levels.

How it works

  1. Discover agents. Defender identifies activity on endpoint devices, and GSA provides network-layer visibility. The session describes these signals as feeding agent discovery.
  2. Review the registry and identity. In the demonstration, the administrator opens the agent registry, filters to local agents, and selects an agent such as Claude Code. The view includes agent details and instances on devices; Entra assigns an Agent ID to discovered agents.
  3. Assess risk and access. Review the agent’s risk, relevant instances, and the data or tools it can access. Runtime risk reflects current behavior, exposure risk concerns vulnerability to exploitation, and blast-radius risk reflects the potential impact of its access.
  4. Set controls for risk levels. The demonstrated policy blocks a high-risk agent and restricts sensitive operations for a medium-risk one. For example, the agent can still perform a read operation while a more sensitive change is denied.
  5. Respond as risk changes. In the demo, Defender blocks malicious content before it reaches the language model, while Purview detects sensitive information. Those events raise the agent’s risk, and policy then restricts or blocks its subsequent actions.

The example also illustrates why separate identities matter: when the agent is blocked, the user can still act directly. The policy targets the agent instance rather than treating the agent’s activity as indistinguishable from the person’s.

Limitations and alternatives

The features described have different rollout stages: Entra Agent ID for local agents and Dynamic Agent Risk are in preview, while Conditional Access dynamic controls are in limited preview. That distinction matters when planning a deployment or relying on these controls for production enforcement.

The discovery workflow depends on endpoint and network visibility from Defender and GSA. The policy examples also show that choosing a response involves tradeoffs: a full block can interrupt an agent’s work, while restricting sensitive operations may preserve lower-risk tasks. Teams should decide which actions count as sensitive and what response fits each risk level.

Dynamic risk combines several signals, but the policy still needs an appropriate response configured for the organization. The demonstration’s medium-risk and high-risk examples show one possible approach: constrain a sensitive operation at medium risk, then block the agent at high risk. They are examples, not a universal policy template.

Sources

Frequently asked questions

What does Microsoft Entra Agent ID do for local agents?

It assigns a distinct identity to discovered local agents so their actions can be separated from the user’s actions and governed with agent-specific policies.

Which local agents are named as supported?

The announcement slide names supported local agents such as GitHub Copilot CLI and Claude Code.

What is the preview status of Dynamic Agent Risk?

The session identifies Dynamic Agent Risk as a preview feature. It combines signals from Entra, Defender, and Purview.

Can Conditional Access block an agent without blocking its user?

Yes. The demonstrated controls target the agent separately, so the user can continue to act directly even when the agent is blocked.