Glossary

AI DLP

AI DLP is data loss prevention applied to the prompt layer: detecting sensitive data in text submitted to AI tools — before submission — rather than in files, email or network traffic. Also called LLM DLP or prompt-layer DLP.

PATTERNS CONTEXT YOUR RULES
Each layer narrows what the next has to look at.
Also called LLM DLP Prompt-layer DLP GenAI DLP

Why a separate category

Traditional DLP is organised around files and channels.

Its model is that data lives in documents and leaves through enumerable exits: email attachment, USB, print, cloud sync, upload.

Text typed into a browser composer is none of those. It never becomes a file, never traverses a monitored channel, and never matches a rule — so a fully functioning endpoint DLP deployment reports nothing while thousands of records flow to a third party every week.

Full comparison

What the system does

Five steps, and the third is where products differ.

The distinguishing property is timing. Inspection happens before the message is sent, which is the only point at which prevention rather than after-the-fact alerting is possible.

  1. 01

    Capture

    The outbound text block and its destination, at the composer.

  2. 02

    Detect

    Sensitive data in it — patterns, contextual models, custom rules.

  3. 03

    Classify

    Against compliance frameworks, scored by whether the destination is sanctioned.

  4. 04

    Enforce

    Log, warn inline, or block the submit.

  5. 05

    Report

    A record to the console and the SIEM.

The two architectures

Where step two executes is the whole evaluation.

Which one a vendor uses is the single most consequential question in an AI DLP evaluation, and it is not always in the datasheet.

Cloud classification

The prompt is sent to a service for inspection. Heavier models, no endpoint footprint — and the content now exists in another vendor’s systems.

On-device classification

Detection runs locally and only a derived, redacted record is transmitted. Harder to build, and the content never leaves the machine.

Which architecture is your vendor?

It is answerable in an afternoon with a packet capture, and most datasheets do not say.