AI agents: access granted. Trust not included | THR877
Leandro Iwase and Ehud Itshaki present an intermediate-level Microsoft Ignite 2026 session on securing AI agent access with Microsoft Entra Agent ID. They cover identity architecture, agent discovery, risk signals, and access controls, followed by demonstrations of policy enforcement.
Separate identities for AI agents
AI agents need access to applications, data, and resources, but inheriting a user’s permissions can give them more access than they should have. The session explains the need for separate agent identities rather than treating user access as a blank check for an agent.
The architecture discussion spans Microsoft and third-party agents, including agents running in the cloud and locally. Microsoft Entra Agent ID provides the focus for securing agent connections and limiting risky permissions.
Discover, manage, govern, and protect agents
The presenters organize agent security around four activities:
- Discover: Identify agents, including those running locally.
- Manage: Manage agent identities and access.
- Govern: Apply controls to agent permissions and resource use.
- Protect: Respond to risk and block compromised access before it becomes an attack path.
The discovery segment covers Microsoft Defender and Global Secure Access, with a later demonstration showing discovery, risk scoring, and visibility into local AI agents. Network controls are also part of the session’s approach to reducing risk.
Dynamic risk and Conditional Access
The session covers unified, dynamic agent risk signals across Microsoft Entra, Microsoft Defender, and Microsoft Purview. It then examines how Conditional Access can use dynamic controls to restrict or block risky agents.
The announcement segment addresses previews for Entra Agent ID, dynamic agent risk, and Conditional Access. The demonstrations show how these controls respond to agent activity, including:
- Enforcing agent policies.
- Blocking malicious prompts.
- Applying data loss prevention (DLP) protections.
- Responding through Conditional Access.
Session chapters
- 00:00: Introduction to securing AI agents with Microsoft Entra identities.
- 00:31: Risks of agents inheriting user permissions and the need for separate identities.
- 01:19: Agent identity architecture across Microsoft, third-party, cloud, and local agents.
- 03:03: Framework for discovering, managing, governing, and protecting agents.
- 04:14: Discovering local agents with Microsoft Defender and Global Secure Access.
- 06:18: Unified dynamic agent risk signals across Entra, Defender, and Purview.
- 08:04: Conditional Access controls for restricting or blocking risky agents.
- 09:45: Preview announcements for Entra Agent ID, dynamic agent risk, and Conditional Access.
- 11:09: Demonstration of agent discovery, risk scoring, and local agent visibility.
- 13:50: Demonstration of policy enforcement, malicious prompt blocking, DLP protections, and Conditional Access responses.