Microsoft adds dynamic controls for enterprise AI agents

Microsoft's October 19 update extends Project Perception and Microsoft Entra Agent ID to identify enterprise agents and adapt their access as behavior and risk change. Microsoft says its discovery coverage is expanding by 35 local agent types, to 60 total, including Claude Code, OpenAI Codex, and GitHub Copilot. It also plans to assign identities to local and previously unknown agents found by security sensors. (Microsoft Security Blog; Microsoft Learn)

Microsoft describes Dynamic Agent Risk as a continuously updated view built from signals across identity, data, endpoints, cloud, network, runtime, and agent activity. Its proposed Dynamic Permissions can limit an agent's authority to a task and time window. Dynamic Conditional Access can adjust that authority as risk changes, such as allowing reads while blocking writes, narrowing accessible resources, requiring verified human approval, or isolating and stopping the runtime. (Microsoft Security Blog)

The update also adds playbooks and Guardian Agents. Microsoft says they can investigate why an agent's risk changed, recommend or apply containment, and use the findings to improve policy. (Microsoft Security Blog)

Microsoft's announcement does not state a product release stage or date. GeekWire reports that the announced features are in private preview and that broader availability is planned for Microsoft Ignite. Microsoft's event page currently lists Ignite for November 17 through 20, 2026. (GeekWire; Microsoft Ignite)

The product design described here links agent identity, task-level authority, and independent enforcement. That is an inference from Microsoft's announcement. The post is a vendor description, not an independent evaluation of detection accuracy or containment effectiveness.