Design Conditional Access policies for agent identities

Short summary

This article explains how to design Microsoft Entra Conditional Access policies for AI agent identities, distinguishing between on-behalf-of agents, autonomous agents, and agent-user identities. It recommends using custom security attributes like approval status, data sensitivity, and environment to target policies rather than creating one policy per agent. A staged rollout using report-only mode, pilot agents, and production rings is advised to safely enforce access controls for non-human identities.

• •Classify agents by access pattern (on-behalf-of, autonomous, agent-user) before designing Conditional Access policies
• •Use custom security attributes (ApprovalStatus, DataSensitivity, Environment) for policy targeting instead of per-agent policies
• •Roll out enforcement in rings starting with report-only mode to validate impact before blocking access

Generated with AI, which can make mistakes.