Microsoft announced a sweeping set of new security controls for enterprise AI agents on October 19, directly addressing the growing concern around autonomous AI systems operating within corporate environments. The new capabilities—Dynamic Permissions, Dynamic Agent Risk, and Dynamic Conditional Access—represent the company’s most comprehensive framework yet for managing agent privileges at scale.
The timing is significant. Just days before, reports emerged of a single hacker using AI agents to breach over 100 companies. Microsoft’s announcement explicitly references this shift in the threat landscape, framing the new controls as essential for enterprises deploying AI agents.
Dynamic Permissions allows organizations to limit agents to approved tasks while adjusting access levels as security risks change—all without revoking the sponsoring employee’s permissions. This addresses a core tension in enterprise agent deployment: agents need enough access to be useful, but excessive privileges create unacceptable risk surfaces.
Dynamic Agent Risk takes this further by continuously assessing agent behavior and adjusting trust levels in real-time. An agent that begins acting outside expected parameters can be automatically constrained or shut down without waiting for human intervention. The system evaluates factors like data access patterns, tool usage, and output characteristics to build risk profiles.
Dynamic Conditional Access ties these capabilities into Microsoft’s broader identity infrastructure through Microsoft Entra, enabling security teams to apply consistent policies across both human and agent identities. This unified approach recognizes that in enterprise environments, agents increasingly act on behalf of employees—creating new categories of “insider” risks.
“Agents are your newest insiders,” Microsoft’s security blog stated explicitly. “Security must move at their speed.”
The October 2026 release marks an architectural inflection point for Microsoft Entra. As organizations transition from managing exclusively human directory objects to securing autonomous AI agents acting on behalf of employees, identity and network perimeter boundaries must fundamentally fuse.
For enterprises deploying Agent 365 and similar systems, these controls offer a path toward responsible agent adoption—providing the visibility and control that security teams require while preserving the productivity benefits that make agents valuable in the first place.