Enable Shadow AI detection for local agents on endpoints
Implementation Effort: Low – Toggle-based enablement in the Shadow AI page of the Microsoft 365 admin center; requires Intune enrollment for managed Windows devices.
User Impact: Low – Detection scans run silently on managed devices; end users are not notified or affected.
Overview
Local AI agents installed directly on employee devices represent an emerging category of shadow AI that operates entirely outside the organization's cloud-based governance controls. Unlike web-based AI applications that can be discovered through network traffic analysis or Defender for Cloud Apps, local agents — such as OpenClaw, GitHub Copilot CLI, and Claude Code — execute on the endpoint itself. They can read local files, execute code, interact with development environments, and act on behalf of the user without ever touching a managed cloud service. This makes them invisible to the cloud-based shadow AI discovery mechanisms that the organization has already deployed for web and SaaS AI applications.
The Shadow AI page in the Microsoft 365 admin center, enabled through the Frontier preview program, provides a dedicated detection surface for these local agents. Administrators enable continuous detection for specific Shadow AI agent types, which triggers scanning of Intune-managed Windows devices to identify endpoints where these agents are installed. The detection results surface a list of devices running each agent, including device name, type, operating system, and last Intune scan timestamp. This gives the security team the same foundational visibility into local agent usage that Defender for Cloud Apps provides for cloud-based AI applications — answering the question of which devices in the organization are running unsanctioned local AI agents and at what scale.
This supports Assume breach by ensuring that local AI agents — which have direct access to files, code, and credentials on the endpoint — are detected before they can be exploited as attack vectors or data exfiltration paths. It supports Verify explicitly by extending the organization's AI agent inventory to include locally-installed agents that would otherwise remain invisible to every other discovery mechanism in the AI governance stack. Without this detection capability, local agents accumulate across the device fleet without any organizational awareness, and the security team cannot assess the scope of unsanctioned local AI usage, let alone make informed decisions about whether to allow or restrict it.