Microsoft previews expanded Entra controls for local AI agents

Microsoft announced expanded Entra controls for AI agents in a November 17 Ignite post, including previews for discovering local agents and adjusting their access as risk changes. The announcement says discovery covers 60 local agent types, including Claude Code, OpenAI Codex and GitHub Copilot CLI, with each discovered agent receiving an identity.
Dynamic Agent Risk combines signals from identity, endpoints, data, networks and agent activity. Microsoft says administrators can distinguish local agents from the employees they assist, potentially blocking a risky agent without stopping the person’s work. These capabilities are previews, not a claim that every agent is automatically governed.
The post also highlights an MCP firewall preview. Technical documentation describes controls for remote servers, individual tools and protocol methods. It requires TLS inspection and does not cover local servers, non-HTTP transports or JSON-RPC batches.
This is an Ignite expansion and roundup, not evidence that every listed component first appeared today: the firewall documentation carries an August update date. The practical direction is more granular control of agents and their tools, with deployment prerequisites and coverage gaps that organizations still need to assess.
Sources and citations
Revision history
Published .