# AI DLP on the endpoint catches pastes and uploads to AI sites well and misses what agents read and send on their own

GUIDE, read 12 Nov 2026

**The short answer.** AI DLP on the endpoint applies data loss prevention to pastes, uploads and agent reads involving AI tools. Microsoft Purview endpoint DLP covers Windows 10/11 and the three latest macOS releases, with browser paste and upload rules. Agents that read files and call tools directly sit outside most classic DLP rules.

## AI DLP is classic DLP pointed at a new set of destinations and a new kind of actor

Data loss prevention (DLP) finds sensitive content and decides whether an action may proceed: allow, warn, block or audit. On the endpoint, the actions DLP knows about are the ones people take: copy a file to a USB drive, print, upload to a website, paste into a page. AI changes two things. The destinations are new, because every employee now has a chat box one tab away. And the actor is sometimes software, because an agent can read a folder, call a tool and send the result to a model without a person touching the keyboard.

This guide separates three flows that vendors often bundle under the label AI DLP: a person pasting text into an AI tool, a person uploading a file to an AI tool, and an agent reading data and sending it on. The first two are covered by mature controls. The third is where coverage thins out, and where a buyer should ask the most questions.

Facts about Microsoft Purview come from Microsoft's documentation, read on 12 Nov 2026. Purview is the best-documented example, and it is not the only product that offers these controls.

## Purview endpoint DLP covers paste and upload to AI sites in the browser

On the AI side, Microsoft's Purview page says DLP policies can warn or block users from sharing sensitive information with third-party generative AI sites accessed through a browser. Its example is a user prevented from pasting credit card numbers into ChatGPT, or shown a warning with an override. The same page says AI apps can be detected through browser activity and categorized as Generative AI in the Defender for Cloud Apps catalog (Microsoft Learn, Purview and generative AI apps page).

For Chrome on Windows 10 and 11, the Microsoft Purview extension for Chrome is the component that lets the organization monitor attempts to access or upload a sensitive item to a cloud service in Chrome and enforce protective actions (Microsoft Learn, Purview extension for Chrome page). So browser-level paste control for Chrome depends on that extension being installed and not removable by the user.

- Upload to a restricted cloud service domain, or access from an unallowed browser. If the user is in an unallowed browser, the upload is blocked and the user is redirected to Microsoft Edge.
- Paste to supported browsers. It detects when a user tries to paste content to a restricted service domain, and it evaluates the content being pasted, regardless of how the source item was classified.
- Copy to clipboard. For a protected file, the policy can audit, block with override or block.
- Browser and domain restrictions through sensitive service domain groups, which let you name which AI sites count as restricted.

## Classic DLP rules assume a person at the keyboard

Most DLP policies were written around a user action with a clear start and end: a file is dragged, a button is pressed, a form is submitted. They trigger on content that matches a sensitive information type, a trainable classifier or a sensitivity label.

Agents break these assumptions in four ways. The first is volume: an agent reading a repository opens thousands of files, and a rule that audits file opens is either silent or unusable. The second is indirection: the agent reads a file, summarizes it in memory and sends the summary to a model provider over an API call, so the artifact that crosses the network is a paraphrase, not the original document. A content match on the original pattern may not fire. The third is channel: many agent calls go from a terminal process or desktop runtime to an API endpoint, never through a browser form. The fourth is attribution: the audit log says a process made a network call, not that an AI agent decided to send Q3 customer data.

OWASP's guidance on excessive agency lists the root causes of agent-driven harm as excessive functionality, excessive permissions and excessive autonomy (OWASP GenAI Security Project, LLM06). Each of the three is a data-loss factor when the agent can read broadly and send freely.

## Agent reads and tool calls need controls at the action, not only at the destination

For an agent, the control point is the action: read this file, run this command, call this tool with these arguments. A DLP-aware product for agents should evaluate the data in the call and the destination of the call, then allow, ask or block before the action completes.

Some of this exists inside the agent tools themselves. Anthropic documents permission rules and sandboxing for Claude Code, and managed settings that an organization can deploy to endpoints through MDM or a managed file (Anthropic, Claude Code settings and server-managed settings pages, read 12 Nov 2026). Those are guardrails on what the agent may touch. They are not content inspection. They answer whether the agent may read a folder, and they do not answer whether the file contains a customer record.

Product claims for content inspection of agent traffic should be read as vendor claims. The questions that test them are in the buyer checks below.

## What each flow needs, and where classic coverage ends

The table compares the three flows on what a mature control looks like today.

| Flow | Typical control | Where it runs | Known gap |
| --- | --- | --- | --- |
| Person pastes text into an AI site | Paste rule with warn or block, evaluated on the pasted content | Browser extension or endpoint DLP component | Only browsers the product supports; other apps and agents are outside |
| Person uploads a file to an AI site | Upload rule on restricted service domains, with an allowed-browser list | Endpoint DLP plus browser component | Desktop AI apps that do not use a managed browser |
| Agent reads files and sends content to a model | Permission rules on file scope, network allowlists, sandboxing | The agent tool's own settings, EDR, or a dedicated agent security product | Content inspection of API traffic is uncommon; paraphrased content evades pattern match |
| Agent calls an MCP tool with sensitive arguments | Server allowlists, tool-call policy | Client policy or a product that sees tool calls | Few products show the arguments; check in trial |

## Labels and classifiers help, if the agent's path touches them

Sensitivity labels travel with files, and a label can drive a block decision on a copy, upload or paste. They help only where the enforcement point sees the label. Microsoft notes that endpoint DLP cannot detect a sensitivity label from another tenant on a document (Microsoft Learn, endpoint DLP page). If agents read files that carry labels from partners or from a second tenant, the label will not protect the document.

Unlabeled data is the larger problem. Source code, internal tickets, customer exports and meeting notes often have no label. Pattern-based types, such as card numbers, catch some of it. A rule for source code or contract text needs a classifier or a keyword set your team maintains. Expect to tune. A rule that fires on every paste of a code snippet will be turned off within a week.

## Warn first on paste, block on upload, and measure before you enforce

A sensible rollout sequence follows from how people actually use AI tools.

- Start in audit mode for 30 days on the AI sites you sanction and the ones you do not. Count pastes and uploads per site and per sensitive type.
- Decide the sanctioned set. For sanctioned sites, allow paste of low-sensitivity content and warn with override on medium. For unsanctioned sites, block high-sensitivity content and offer a link to the approved alternative.
- Block uploads of labeled or classified files to unsanctioned AI sites, and set an allowed-browser list so uploads cannot be made from a browser your policy does not reach.
- Review override reasons weekly in the first month. They show which rules are wrong and which workflows lack an approved path.
- Write the agent rules separately: file scope per project, no read of credential stores, a network allowlist, and an approval step for outbound calls to new domains. Enforce these in the agent tool's managed settings where available.

## Questions that test a vendor's AI DLP claim

Ask each question in a trial, using test data you control.

Which browsers and which operating systems does the paste rule run in, and what happens in an unsupported browser? Microsoft's own table shows paste support varying by platform, with some entries marked preview, so ask for the current matrix.

What does the product see from a desktop AI app that is not a browser? Paste a test string into it and look for the event.

What does the product see when a command-line agent reads a file and calls a model API? Ask to see the log line, and ask whether the content of the call is inspected or only the destination.

Can the product stop the action before it completes, or does it alert after the data has left? Ask for a demonstration of a block on your machine.

What can a user do to disable the control? Try it.

How are overrides recorded, and who sees them?

A vendor's answer is a vendor claim. On this site's scorecard a claim stays capped at 7 of 10 until a second source confirms it, and a feature with no evidence found scores 0.

## Where this sits in the index

- [Agent Surface Map](/agent-surface-map/): Which of six agent surfaces each of 16 tools sees
- [Action Control Matrix](/action-control-matrix/): Which of six agent actions each tool can stop
- [The ranked index](/ranking/): 16 tools on seven published criteria

## Related guides

- [Shadow AI detection signals](/guides/shadow-ai-detection-signals/)
- [Agent surfaces explained](/guides/agent-surfaces-explained/)
- [Browser agents and AI extensions](/guides/browser-agents-and-ai-extensions/)

## Sources

- Microsoft Learn: endpoint DLP: https://learn.microsoft.com/en-us/purview/endpoint-dlp-learn-about (read 12 Nov 2026)
- Microsoft Learn: Purview and generative AI apps: https://learn.microsoft.com/en-us/purview/ai-microsoft-purview (read 12 Nov 2026)
- Microsoft Learn: Purview extension for Chrome: https://learn.microsoft.com/en-us/purview/dlp-chrome-learn-about (read 12 Nov 2026)
- OWASP GenAI: LLM06 excessive agency: https://genai.owasp.org/llmrisk/llm062025-excessive-agency/ (read 12 Nov 2026)
- Anthropic: Claude Code settings: https://code.claude.com/docs/en/settings (read 12 Nov 2026)
- Anthropic: Claude Code server-managed settings: https://code.claude.com/docs/en/server-managed-settings (read 12 Nov 2026)

Last reviewed 12 Nov 2026. Canonical: https://aiendpointsecurityranked.com/guides/ai-dlp-on-the-endpoint/
