Microsoft 365 Copilot can only use content the signed-in user can already open, so the real Copilot risk is content that too many people can open. Microsoft Purview's data security posture management for AI (DSPM for AI) brings the tools for that problem together: insight into AI activity, data risk assessments that find overshared SharePoint content, and one-click policies that switch on detection and protection. In this post I'll cover where the feature lives now, the prerequisites, how to run and read an assessment, which one-click policies to turn on, and what to fix first.
Status (October 2026): The original experience is now listed in the Purview portal as DSPM for AI (classic). Microsoft has replaced it with the current Data Security Posture Management (Solutions › DSPM), which covers AI apps and agents alongside the rest of your data, and says most new features will be added only there. Microsoft's documentation calls the oversharing scans data risk assessments, and the one-click policies it documents are named with a "DSPM for AI" prefix. Re-check Microsoft Learn before you plan, because this area changes quickly.
Prerequisites and how it works#
Open the Microsoft Purview portal (purview.microsoft.com) and go to Solutions › DSPM. Don't confuse it with DSPM (classic) or DSPM for AI (classic), which are still available. You need permissions for compliance management, such as the Microsoft Entra Compliance Administrator role or membership of the Purview Compliance Administrator role group; the Security Copilot features also need the Data Security Viewer role. If you use administrative units, only an unrestricted administrator can create the one-click policies, because they apply to all users. Restricted administrators see policy results, reports and activity only for users in their unit. My guide to Purview roles and role groups covers the permission model.
| To see or protect | You need |
|---|---|
| Microsoft 365 Copilot and agent interactions | Purview auditing turned on (the default for new tenants) and users licensed for Microsoft 365 Copilot |
| Third-party AI sites such as ChatGPT or Gemini | Devices onboarded to Microsoft Purview, plus the Microsoft Purview browser extension on Windows for Chrome |
| Other AI apps in Microsoft Edge | An Edge configuration policy that activates the Purview integration in Edge |
| AI apps other than Microsoft 365 Copilot and Microsoft Facilitator | Pay-as-you-go billing set up for your organization |
Check auditing in Exchange Online PowerShell before anything else; True is what you want:
Connect-ExchangeOnline
Get-AdminAuditLogConfig | Format-List UnifiedAuditLogIngestionEnabledFor the endpoint pieces, see Endpoint DLP onboarding and deploying the Purview browser extension.
Step-by-step#
1. Accept the setup tasks and pick the objectives#
The first time you open DSPM, you're prompted to accept the initial setup tasks that aren't already on in your tenant. Allow a day or so before tenant data appears. After that, the Objectives page shows security goals as cards, each with a remediation plan of one-click policies and recommended actions. For Copilot, the two that matter are Prevent data exposure in Microsoft 365 Copilot and Microsoft Copilot interactions and Prevent oversharing of sensitive data. AI observability lists the AI apps and agents with activity in the last 30 days, how many are high risk and how many had sensitive interactions.
2. Read the default data risk assessment#
Go to Discover › Data risk assessments. With no setup, Purview assesses the top 100 SharePoint sites by usage every week; the first results take about four days after the default assessment is created. The summary shows total items, sensitive items detected and links that share data with anyone. Select View details, then a site, to open a pane with four tabs:
- Overview: the site's summary.
- Identify: how much content has and hasn't been scanned for sensitive information types, with an on-demand classification scan.
- Protect: the remediation options in step 4.
- Monitor: items shared with anyone, with everyone in the organization, with specific people and externally.
Use Export (Excel, CSV, JSON or TSV) to hand the list to site owners.
3. Run a custom assessment for your Copilot pilot#
Select Create custom assessment, choose all users or just your pilot group, and pick the sites to scan. Wait until the status shows Completed, at least 48 hours. Custom results don't refresh and expire after 30 days, so use Duplicate to rerun the same scope later. For file-level detail, choose Item-level on the Scan level page. It needs an Entra app registration with these Microsoft Graph application permissions and admin consent: Application.Read.All, Directory.Read.All, Files.ReadWrite.All, SensitivityLabels.Read.All, Sites.ReadWrite.All and User.Read.All. Item-level scanning covers SharePoint only (not OneDrive), up to 10 sites, and flags items with sharing links for external or anonymous users. For each item you can Resolve, Apply sensitivity label, Notify the site owner or Remove sharing link.
4. Remediate from the Protect tab#
| Option | What it does | Use it for |
|---|---|---|
| Restrict access by label | Creates a DLP policy that stops Microsoft 365 Copilot and agents summarizing items with the labels you choose | Labeled crown-jewel content (see Purview DLP for Copilot) |
| Restrict all items | Uses SharePoint Restricted Content Discovery to exempt the site from Copilot | Sensitive sites that need time to clean up |
| Create an auto-labeling policy | Labels unlabeled files that contain sensitive information | Sites full of unlabeled sensitive data (see service-side auto-labeling) |
| Create retention policies | Deletes content that hasn't been accessed for at least three years | Stale content nobody needs |
The Monitor tab also links to a SharePoint site access review, which hands the fix to site owners; the workflow is in finding and fixing SharePoint oversharing.
5. Turn on one-click policies deliberately#
One-click policies are created in the owning solution, and you edit them there. The ones most relevant to Copilot:
| Policy | Solution | Effect |
|---|---|---|
| DSPM for AI - Protect sensitive data from Copilot processing | DLP | Blocks Microsoft 365 Copilot and agents from processing items with the selected labels |
| DSPM for AI - Detect risky AI usage | Insider Risk Management | Feeds risky prompts and responses into user risk |
| DSPM for AI - Unethical behavior in AI apps | Communication Compliance | Detects sensitive information in prompts and responses, for all users |
| DSPM for AI: Detect sensitive info added to AI sites | DLP (audit only) | Discovers sensitive content pasted or uploaded to AI sites in Edge, Chrome and Firefox |
| DSPM for AI - Block sensitive info from AI sites | DLP (test mode) | Block with override for elevated-risk users, using Adaptive Protection |
Policies that use Adaptive Protection turn it on with default risk levels if it isn't on already; read how Adaptive Protection works first.
Verify#
- Reports populate after at least a day; filter by Copilot experiences and agents, Enterprise AI apps and Other AI apps.
- In Discover › Activity explorer, the AI activities tab should show AI interaction, Sensitive info types and, once DLP is enforcing, DLP rule match events. Prompts and responses appear only if you have the right permissions.
- Test with a pilot user: open a file with test sensitive data through Copilot, then look for the event.
- After remediation, duplicate the custom assessment and compare counts; watch the weekly default assessment trend down.
Tips & gotchas#
- Fix in this order: sites holding sensitive data with Anyone or organization-wide links; then unlabeled sensitive content; then labeled content Copilot shouldn't summarize; then stale content.
- Remove sharing link can cut off legitimate users. Use it sparingly and ask owners to replace the link with a narrower one.
- Assessments are capped at 200,000 items per location, and file counts may be inaccurate above 100,000 files per location.
- If you delete a one-click policy, the Policies page shows
PendingDeletionuntil it's gone. Policies created in the early "Microsoft Purview AI Hub" preview keep their old name prefix. - The AI interaction event shows no prompt or response for users without an Exchange Online mailbox.
