Skip to main content

Deploy Endpoint and Browser DLP Policies for AI Data Locations

Implementation Effort: Medium – Requires configuring endpoint DLP policies for onboarded devices, deploying browser DLP for AI interactions in Microsoft Edge, and optionally leveraging one-click policy deployment from Data Security Posture Management (DSPM) to accelerate coverage.
User Impact: Medium – Users on onboarded devices will experience restrictions when attempting to copy, paste, or upload sensitive content to AI sites; browser-level DLP in Edge will block or warn when sensitive data flows through AI interactions in the browser.

Overview​

Server-side DLP policies configured at the Microsoft 365 Copilot and Agent 365 locations prevent AI services from processing highly classified content. However, those policies only control what happens within the AI service itself — they do not govern what happens on the user's device or in the browser after content has been retrieved. Endpoint DLP and browser DLP extend data loss prevention enforcement to the client side, closing the gap between server-side policy and client-side behavior.

Endpoint DLP policies restrict what users can do with sensitive content on their onboarded devices. When a user works with a sensitive file on the device, endpoint DLP can block copy-to-clipboard actions, restrict printing, prevent saving to unauthorized locations, and control upload to cloud services, all based on the sensitivity label or sensitive information type present in the content. For AI scenarios, this matters because even when server-side DLP prevents Copilot from including highly classified content in a response, users may still access those same documents directly and attempt to move the content through client-side actions. Endpoint DLP ensures that the protection follows the data to the device level.

DSPM provides one-click DLP policies that accelerate endpoint and browser DLP coverage for AI apps. Rather than building each policy manually, administrators can create preconfigured policies that, for example, give users at elevated risk a block with override when they paste or upload sensitive information to AI apps in Edge, Chrome, and Firefox, or block prompts that contain common sensitive information types from being sent to AI apps in Microsoft Edge. Because the one-click policies apply to all users, administrators should review the scope and mode of each policy after it is created.

Browser DLP for AI interactions in Microsoft Edge adds a further enforcement layer. When users interact with AI services through Edge — including Copilot in the browser, enterprise AI apps, and web-based AI tools — browser DLP policies can detect and block sensitive data in prompts before it is submitted to the AI service. This is a different threat vector than the server-side scenario: rather than preventing AI from surfacing sensitive content, browser DLP prevents users from inputting sensitive content into AI services. This is particularly relevant for AI apps that are not covered by the Microsoft 365 Copilot DLP location, where browser DLP adds an enforcement point at the browser boundary for the apps that Microsoft Purview supports in Edge for Business.

This activity supports Assume Breach by enforcing data protection controls across multiple layers — server, endpoint, and browser — so that a failure or bypass at any single layer does not result in uncontrolled data exposure. It supports Use Least Privilege Access by restricting client-side actions on sensitive content to only what is authorized by policy, regardless of whether the user accessed the content through an AI interaction or directly. Without client-side DLP enforcement, server-side DLP policies create a false sense of security — highly classified content is blocked from Copilot responses, but the same content can be freely copied, printed, or uploaded from the user's device through other channels. Threat actors who compromise an endpoint can exfiltrate AI-adjacent sensitive content through client-side actions that server-side policies cannot see.

Reference​