A laptop security interface scans a local AI agent’s activity and blocks a tool call involving sensitive customer data.
Microsoft wants data loss prevention to cover AI agents that run on users' own computers, not just AI in the browser. A new Microsoft 365 roadmap entry, ID 577862, says Microsoft Purview Endpoint DLP will extend inline protection to supported local AI agents. The goal is to stop sensitive data from leaking while people talk to agents and while agents call tools.

This is not available yet. The roadmap lists the feature as in development, with preview planned for January 2027 and general availability for March 2027. For admins already dealing with unapproved AI coding assistants on company laptops, it is still worth planning for now.

A laptop security interface scans a local AI agent’s activity and blocks a tool call involving sensitive customer data. What the roadmap entry says​

Here is what the listing states:

FieldValue
Roadmap ID577862
ProductMicrosoft Purview
StatusIn development
Preview (estimated)January 2027
General availability (estimated)March 2027
Release ringsPreview, General Availability
Cloud instanceWorldwide (Standard Multi-Tenant)
Platform listedWeb

According to the description, organizations will be able to apply DLP policies to prompts, files, and tool interactions. Those policies can block data movement that breaks policy "at supported interaction points." Microsoft says the controls are meant to prevent sensitive data exposure as agents access company information and act on users' behalf.

Microsoft's own roadmap disclaimer applies here. The roadmap provides estimated release dates and descriptions for commercial features, and all of that information is subject to change. January and March 2027 are estimates, not promises.

Summary: Purview DLP is planned to reach local AI agents, with preview targeted for January 2027 and GA for March 2027. "Supported" is the key qualifier throughout.

Microsoft's Ignite announcement adds detail​

The roadmap entry is brief, but Microsoft's Tech Community security blog announced the same capability at Ignite with more detail. The post lists "Protect the sensitive data local agents use" as a public preview January 2027 item. It says admins will gain visibility into local agent activity and apply DLP policies that prevent local agents from processing sensitive data, whether it comes from a prompt or from the files and tools the agent accesses.

The same post says admins will be able to protect sensitive data at runtime by blocking prompts and tool calls that contain sensitive information in text, files, and links. It uses a local AI coding agent as its example and describes how a Purview DLP policy scoped specifically for Claude Code agents can identify the sensitive content and block the interaction before the data is exposed at runtime.

This is the main point for Windows admins. Earlier endpoint DLP looked at what a person did, such as pasting, uploading or copying to USB. This feature also covers what an agent does when it reads a file, calls a tool or sends a prompt to a model.

This builds on the Build 2026 preview​

Microsoft started on local agents earlier this year. At Build, Microsoft's Purview team announced it was extending Purview visibility and protection capabilities to local agents developed on GitHub Copilot CLI, Claude Code, OpenAI Codex, and OpenClaw - in Public Preview. That work focused heavily on observability: visibility on Purview Data Security Posture Management (DSPM) across agent inventory, as well as into how local agents interact with sensitive data—across prompts, responses, and actions.

Read together, roadmap item 577862 looks like the move from watching local agents to enforcing policy on them inline at runtime.

One caution: the roadmap entry does not name any supported agents. Claude Code appears in Microsoft's Ignite example, and the Build preview named four agent platforms. Neither source guarantees that every one of those agents, or every action they take, will be covered at preview or GA.

Why browser DLP isn't enough​

Purview can already do a fair amount with AI, so it helps to be clear about what's new.

Current Endpoint DLP for generative AI sites. Microsoft Learn documents that Windows computers that are onboarded to Microsoft Purview can be configured for Endpoint data loss prevention (DLP) policies that warn or block users from sharing sensitive information with third-party generative AI sites that are accessed via a browser. The standard example: a user is prevented from pasting credit card numbers into ChatGPT, or they see a warning that they can override. In that case a person is using a website.

Current Agent 365 DLP. For Agent 365 agent instances, DLP is supported by explicitly specifying agent instances in the DLP policy as you would a user, or by specifying a security group that includes agent instances. The documented coverage is limited to block or audit agent-to-human and human-to-agent for Microsoft Teams, OneDrive or SharePoint, and emails. Agent-to-tool blocking is not on that list.

Entra-registered custom AI apps. These follow a separate, opt-in path. Industry coverage summarizing Microsoft Learn describes DLP for them as requiring a PowerShell cmdlet scoped to that Entra-registered app, plus developer integration with the Purview APIs. The app has to cooperate.

Local agents fall outside all three. A coding agent in a terminal reads files directly from disk, runs tools and talks to model endpoints, often with no browser involved. In a review of Purview DLP, the vendor Nightfall observed that purely local stdio MCP communications do not traverse a network inspection point. Nightfall sells a competing product, so weigh that framing accordingly, but the technical point holds. Network and browser controls can't see a process on the laptop talking to another process on the same laptop. That gap is what this roadmap item is meant to close.

Summary: Browser DLP covers people on AI websites, Agent 365 DLP covers specific Microsoft 365 channels, and custom apps need developer integration. Roadmap 577862 targets local agents and their tool calls, which none of those cover.

Where it fits in Microsoft's agent security stack​

Purview is one layer of Microsoft's broader plan for governing agents on Windows. WindowsForum's coverage of Microsoft's enterprise AI agent rollout plan describes a layered model:

  • Microsoft Entra Agent ID handles identity.
  • MXC handles execution containment.
  • Intune distributes policy.
  • Defender handles threat detection.
  • Purview provides Endpoint DLP for supported agent prompts, tool calls and tool output.

The guidance there is practical: start with a test policy, for example one that blocks a test sensitive-information pattern for the target agent. Then wait for the policy to sync, confirm it is in effect on the endpoint, then test both pasting data directly into a prompt and having the agent read a test file.

Testing both paths matters because an agent can expose data either way. A user can paste a customer list into a prompt, or the agent can open customers.xlsx itself because it decided that would help.

A caveat from Microsoft's own documentation​

The Agent 365 documentation contains a warning that likely applies to agent DLP generally: because an agent instance is unaware of the block action, the agent owner must actively monitor a DLP policy that uses this configuration and understand the impact to subsequent workflows.

A blocked human sees a policy tip and stops. A blocked agent may not know why its tool call failed. It could retry, try a different approach, or finish its task with missing data and report success. Microsoft hasn't said whether local-agent enforcement will behave this way. Admins should still assume blocks will interrupt agent workflows in ways that are hard to notice, and plan their monitoring for that.

What we still don't know​

The roadmap and the Ignite post leave several important questions open:

  • Which agents are "supported." The roadmap does not list them.
  • Which interaction points are enforced. "Supported interaction points" means coverage will not include every agent action.
  • OS coverage. Endpoint DLP in general extends the activity detection and protection capabilities of DLP to sensitive items that are physically stored on Windows 10, Windows 11, and macOS (Catalina 10.15 and higher) devices. The roadmap entry lists only "Web" as its platform and does not confirm which endpoint operating systems the new controls will support.
  • Licensing. No licensing detail has been published for this feature.
  • Configuration steps. Microsoft has not documented how to set it up.

Anyone who tells you exactly how this will work in production before January is guessing.

What admins can do now​

Until preview arrives, these steps will help, based on general industry practice rather than a published Microsoft procedure for this feature:

  1. Inventory local agents. Find out which coding assistants and agent CLIs are already installed on managed devices. Expect more than you hoped for.
  2. Confirm device onboarding. Endpoint DLP requires devices onboarded to Purview. If your fleet isn't onboarded, start now.
  3. Review your sensitive information types and labels. Agent DLP depends on your classification. If Purview can't recognize your sensitive data, it can't block an agent from using it.
  4. Map data paths. Note which repositories, file shares and tools agents can reach. The roadmap specifically names agent-to-tool interactions.
  5. Plan for monitoring. Agents may not report blocks, so decide who will watch DLP alerts and Activity Explorer once the preview starts.
  6. Pilot in audit mode first. When preview arrives, audit before you block, or you may stop your developers' agents in the middle of a task without warning.

Bottom line​

Roadmap item 577862 targets a real gap. Agents act faster than users and can read files and call tools on their own, while most DLP still assumes a person is doing the copying and pasting. Microsoft's Ignite announcement confirms the intent: block prompts and tool calls containing sensitive text, files and links, with a January 2027 public preview.

Until Microsoft publishes the list of supported agents, the enforcement points and licensing, treat this as a direction rather than finished protection. Use the next two months to inventory agents, onboard devices and clean up classification so you're ready to test when the preview opens.

 

References

  1. Microsoft Purview: Endpoint Data Loss Prevention-Protect sensitive data across agent-to-human and agent-to-tool interactions with expanded DLP controls across local agents Microsoft 365 Roadmap 2026-11-17T19:05:12.369378Z
  2. Microsoft Purview enables developers with strong data security across AI apps and agents techcommunity.microsoft.com
  3. Microsoft Purview DLP Reviews 2026 nightfall.ai