AI for M365 Admins — Chapter 5: Purview DSPM for AI & Shadow AI Governance (Local Agents, Endpoint DLP & Intune Edge Controls)

📚 AI for Microsoft 365 Administrators Chapter 5 of 10
View Full Book TOC →
🤖 Chapter 5 · Learning Objectives AI for M365 Admins · Purview DSPM for AI & Shadow AI
What You Will Master in Chapter 5
  • Shadow AI is bigger than Copilot: Even if you haven't rolled out Microsoft 365 Copilot, employees are already pasting source code and spreadsheets into consumer ChatGPT, Gemini, Claude, and local coding agents.
  • Purview DSPM for AI: Use Microsoft Purview's unified posture dashboard to track sensitive info types (SITs) entering both Microsoft Copilot and third-party AI web apps.
  • Local Agent Discovery (Ignite 2026): Pair Microsoft Defender for Endpoint + Entra Global Secure Access to discover up to 60 local developer agents (such as GitHub Copilot CLI and Claude Code) running on corporate workstations.
  • Intune + Edge Guardrails: Enforce Endpoint DLP browser restrictions and Intune Edge policies to block unauthenticated consumer AI uploads while steering users to sanctioned Enterprise Copilot Chat.
Connect-IPPSSession; Get-DlpCompliancePolicy | Where-Object { $_.EndpointDlpLocation -ne $null } | Select Name, Mode

In Chapters 1 through 4, we focused on governing Microsoft 365 Copilot inside your tenant boundary. But ask any CISO what keeps them awake at night in 2026, and the answer is Shadow AI:

  • A finance analyst pasting an unreleased Q4 earnings table into a personal ChatGPT or Claude browser tab.
  • A developer running an unsanctioned local coding agent (or an OS-level screen-capturing agent) on a corporate laptop with full read access to local Git repos and browser sessions.

You cannot solve Shadow AI with an Acceptable Use Policy PDF. You need visibility (Microsoft Purview DSPM for AI + Defender local agent discovery) and technical enforcement (Endpoint DLP + Intune Edge policies).


3-Layer Enterprise Shadow AI Defense Architecture Layer 1: Visibility (DSPM for AI) Purview DSPM for AI Dashboard • Tracks M365 Copilot + third-party AI web apps (ChatGPT, Claude, etc.) • Defender + GSA discovers up to 60 local CLI / IDE agents Find Shadow AI Before Blocking Layer 2: Browser & Endpoint DLP Purview Endpoint DLP + Edge • Blocks paste/upload of labeled or SIT-matched files to AI domains • Sensitive Service Domain Groups (chatgpt.com, claude.ai, etc.) Stops Exfiltration in Real Time Layer 3: Intune & Safe Harbor Intune Endpoint + Edge Policies • Enforce Edge Work Profile for Enterprise Copilot Chat • Restrict unsanctioned browser extensions & screen-capture hooks Steers Users to Sanctioned AI Blocking Shadow AI fails unless you simultaneously provide a sanctioned, safe-harbor Enterprise AI workspace.

1. Enabling Purview DSPM for AI One-Click Policies

In the Microsoft Purview portal → Solutions → DSPM for AI:

  1. Verify Prerequisites: Ensure Microsoft Purview Audit is enabled and Windows/macOS endpoints are onboarded to Microsoft Defender for Endpoint / Purview Device Onboarding.
  2. Activate One-Click Policies: Enable Detect sensitive info added to AI sites (monitors clipboard paste and file uploads containing PII/PCI/source code to top generative AI web domains in Microsoft Edge and Chrome via the Purview extension) and Insider Risk Management — Risky AI usage (correlates high-volume prompt activity with resignation dates or mass SharePoint downloads).
  3. Review Activity Explorer: Filter by Activity: Uploaded to cloud / Pasted to browser to see which departments are using unsanctioned AI tools—and why.

2. Blocking Sensitive Uploads to Unsanctioned AI Sites with Endpoint DLP

Rather than blanket-blocking chatgpt.com or claude.ai (which drives employees onto personal phones), Purview Endpoint DLP lets users visit AI sites for generic research while blocking them from pasting or uploading sensitive corporate files:

  1. In Purview → Settings → Data loss prevention → Endpoint DLP settings → Service domains, create a Sensitive Service Domain Group named Unsanctioned-Consumer-AI-Sites containing chatgpt.com, chat.openai.com, claude.ai, gemini.google.com, perplexity.ai, and deepseek.com.
  2. Target your Endpoint DLP Policy to Devices (EndpointDlpLocation) with rules matching Confidential / Highly Confidential sensitivity labels or corporate Sensitive Information Types (SITs).
  3. Under Actions → Upload to a restricted cloud service domain or access from an unallowed browser, set Unsanctioned-Consumer-AI-Sites to Block (or Block with override during a 2-week user education phase).

Auditing Endpoint DLP Policies via PowerShell

POWERSHELL
Connect-IPPSSession -UserPrincipalName [email protected]

# Inspect all active DLP policies protecting Endpoint Devices
Get-DlpCompliancePolicy |
    Where-Object { $_.EndpointDlpLocation -ne $null } |
    Select-Object Name, Mode, Enabled, Priority, WhenChanged |
    Format-Table -AutoSize

3. Discovering Local Coding Agents & Governing Edge via Intune

With the rise of local autonomous agents (GitHub Copilot CLI, Claude Code, Codex, and OS-level screen-reading assistants), browser-only DLP is no longer enough:

  1. Defender + Entra Global Secure Access Local Agent Discovery: Enable endpoint discovery in Microsoft Defender XDR and Entra Agent ID to inventory up to 60 local agent binaries running across Windows and macOS workstations—and require Microsoft Execution Containers (MXC) policies (covered in Chapter 7) for approved developer agents.
  2. Intune Settings Catalog Policies for Microsoft Edge: Deploy a dedicated Intune Configuration Profile (Windows 10/11 → Settings Catalog → Microsoft Edge) with two guardrails:

* Configure Copilot in Edge: Enable Microsoft 365 Copilot Chat with Enterprise Data Protection (signed into Entra ID work profile) and disable consumer Bing/personal sidebar sign-ins. * Extension Allowlist: Configure ExtensionInstallBlocklist = * paired with ExtensionInstallAllowlist to block unverified third-party "AI summarizer" Chrome/Edge extensions that silently read DOM contents of Outlook Web and SharePoint pages.


4. Chapter 5 Lab Checklist

  • [ ] Open Purview → DSPM for AI and verify endpoint telemetry is flowing from managed Windows/macOS devices.
  • [ ] Create the Unsanctioned-Consumer-AI-Sites service domain group in Purview Endpoint DLP settings.
  • [ ] Deploy an Endpoint DLP rule in Audit / Test mode for 14 days to measure real Shadow AI usage before switching to Block.
  • [ ] Configure Intune Settings Catalog for Microsoft Edge to enforce work-profile Enterprise Copilot Chat and block unapproved AI browser extensions.

💬 Reader Comments & Feedback

Did this runbook help, or did Microsoft change a cmdlet or UI menu? Share your feedback below.