L11. Securing Microsoft Copilot and AI Apps with Purview Data Security Posture Management

The Oversharing Problem

Microsoft 365 Copilot answers questions using whatever content the *signed-in user* already has permission to access across SharePoint, OneDrive, and Teams. This is by design (Copilot doesn't grant new access), but it also means years of accumulated oversharing in SharePoint, sites shared "Everyone except external users," stale permissions never cleaned up, forgotten broad links, suddenly becomes instantly discoverable through a single natural-language prompt instead of requiring someone to know where to look.

Identifying Overexposure in SharePoint

Before AI rollout, security teams need visibility into exactly how much content is overshared. Native SharePoint Advanced Management and Purview reporting surface:

• Sites with unusually broad sharing (organization-wide or "anyone" links)
• Sensitive-labeled content sitting in broadly-shared locations
• Stale sites with permissions that no longer match active team membership

This overexposure assessment is a prerequisite step, not an optional one, since Copilot's answers are a direct function of exactly this permission surface.

Microsoft Purview Data Security Posture Management (DSPM) for AI

Purview DSPM for AI is purpose-built to monitor AI usage risk, going beyond a one-time SharePoint audit:

• Discovers interactions with Copilot and other AI apps (including third-party generative AI tools accessed through the browser) across the tenant
• Flags when a sensitive information type or sensitivity label was included in a prompt or a Copilot response, surfacing exactly which users and which content are driving the risk
• Provides recommendations to reduce oversharing risk, directly actionable back into SharePoint permission cleanup
• Extends coverage to custom Copilot Studio agents and Foundry-built AI applications, not just Microsoft 365 Copilot itself, giving one posture view across however many separate AI surfaces an organization has adopted

Real-Time Protection for Copilot Studio Agents

For Microsoft Copilot Studio (the low-code platform for building custom conversational agents), real-time protection applies Purview Data Loss Prevention (DLP) policies directly to agent interactions as they happen:

• A DLP policy can detect a sensitive information type appearing in an agent's generated response and block or redact it before the response reaches the user
• This closes a specific gap: a well-intentioned agent, built quickly by a business user rather than IT, could otherwise surface sensitive data it was never explicitly authorized to expose, simply because its underlying data source contained it

How This Feeds Access Decisions

Sensitivity labels and DLP aren't just reporting artifacts, they are inputs into what Copilot (and DSPM) treats as risky. Content correctly labeled as Highly Confidential is exactly the content DSPM for AI prioritizes when scoring oversharing risk, which is why a mature Purview labeling and DLP program is a prerequisite for a low-risk Copilot rollout, not a parallel, unrelated compliance project. Exam tip: SC-500 frames AI security as fundamentally a data security posture problem first: the exam expects you to recognize that fixing Copilot oversharing risk means fixing SharePoint permissions and labeling, and that Purview DSPM for AI is the tool that specifically surfaces which users, prompts, and content are driving that risk across every AI surface in the tenant, not just Microsoft 365 Copilot.