跳到主要内容

Discover local AI agents with Microsoft Defender for Endpoint

Implementation Effort: Medium - Requires suitable device onboarding, supported Defender versions, and access to the agent inventory.
User Impact: Low - Discovery collects inventory information; it does not block agents or change their access.

Overview​

Microsoft Defender for Endpoint discovers supported local AI agents on Windows and macOS devices, including their configured Model Context Protocol (MCP) servers where supported. The inventory connects installations to device and account context that a cloud-only agent inventory can miss.

This supports Verify explicitly by making local installations visible and Assume breach by identifying unexpected agents and tools. It complements Intune-based Shadow AI discovery and the Agent 365 registry rather than replacing them.

Prerequisites​

  • Onboarded devices in a supported commercial-cloud environment, with the required Defender versions and licensing.
  • A running, reporting Defender client and an agent/platform combination covered by the support matrix.
  • Security data basics (Read) permission with Security for AI as a data source in Defender unified role-based access control (RBAC).
  • Endpoint ownership for reporting gaps and unsupported installations.

Limitations​

  • Process-based discovery detects a running process; it does not require successful sign-in or a prompt, and does not prove successful service use.
  • Some supported VS Code extensions can be inventoried without running the extension or using active-mode antivirus. Process-based discovery and MCP configuration collection still require active-mode Defender Antivirus and real-time protection.
  • An agent can have multiple inventory entries across devices and accounts. A missing entry is not proof that an agent is absent.
  • A configured MCP server is not evidence that a tool was invoked.
  • Discovery does not establish native Entra Agent ID, runtime protection, approval, or least-privilege access.

Reference​