Data loss prevention for AI tools

Data loss prevention (DLP) for AI is the set of controls that stop sensitive information leaving your organisation through an AI tool, sanctioned or not. The discipline isn’t new. What’s new is the volume of low-friction ways to paste data into something external. Controls that assumed file transfers and email now need to account for a chat box.

Start with what you’re protecting

DLP fails when it tries to protect everything equally. Classify first. Decide which data actually matters (customer records, credentials, unreleased financials, regulated personal data) and concentrate effort there. A control that blocks everything gets switched off or worked around, which is the shadow-AI cycle again.

Layers that hold up

Sensitivity labels that travel with the data come first, because they let downstream tools enforce policy instead of guessing. On top of that sits tenant-level policy for the sanctioned tools: configure Copilot so labelled content isn’t summarised, surfaced, or sent to connectors that shouldn’t see it.

Egress controls at the network and endpoint can flag or block sensitive content heading to known consumer-AI endpoints. Imperfect, but they raise the floor. They only work politically, though, when paired with a sanctioned alternative worth using: managed Copilot, or a local model. Enforcement without a good internal option just pushes staff toward creative workarounds.

Last, logging and review. You can’t improve what you can’t see, and audit trails on the sanctioned tools are what turn incidents into lessons.

Good DLP is a ratchet, not a wall: classify what matters, give people a safe route, and make the unsafe route visible and inconvenient.

Treat it as ongoing

AI tools, plugins, and endpoints change monthly. A DLP posture set once and forgotten drifts out of date fast. Review the list of sanctioned tools, connectors, and blocked endpoints on a schedule, not only after an incident.