In short
- Coverage first, then scenarios on production machines, then the event in the SIEM.
Testing endpoint DLP should start with the asset list, not the rule set. Every rule test assumes the agent is running on that laptop with the current policy, and a machine rebuilt from an old image breaks that assumption without raising a single alert.
Endpoint DLP testing means sending decoys off real machines through the channels the agent watches (USB, printing, clipboard, browser, cloud sync) and recording what it blocks or reports.
Diagram · the four-step method
How endpoint DLP fails without an alert
The missing agent tops the list: a reinstall that skipped the onboarding package, or a service that died after an update. Then policies that have not synced for weeks, rules left in audit-only mode after the pilot, exclusions added “temporarily” for a business team, and channels nobody wrote a rule for, such as network printing.
Like other DLP false negatives, these failures vary from machine to machine, so one reference laptop tells you little. Sample: at least one machine per system image, operating system and agent version, one account per excluded group, and the machines with the oldest check-in.
Step 1: map coverage before testing
Cross-check three sources that almost never agree.
- The asset inventory (directory, workstation management tool, CMDB).
- The DLP console: agent version, last contact and policy version for each machine. In Microsoft Purview, the onboarded devices list shows each device’s policy sync status; export it. Start with a seven-day “last seen” threshold, tuned to how often you publish policy changes; a machine past it may be enforcing old rules, so treat it as unprotected until checked.
- The EDR console or antivirus, often more up to date on the list of active machines.
Unmanaged devices
Unmanaged devices (BYOD, contractors, lab machines, developer boxes outside the domain) will never have an agent. Check that they cannot reach sensitive data, or that another control takes over: conditional access, proxy, CASB and SaaS DLP, egress filtering. An unmanaged laptop that reaches SharePoint from a browser is a coverage hole.
Machines people forget
Add macOS and Linux laptops, jump servers, VDI and admin sessions; the agent often covers fewer channels there than on Windows.
Step 2: design the scenarios by channel
A good endpoint scenario fits on one line: who, which machine, what data, which channel, what you expect. Start with everyday actions (copy, paste, upload); clever bypasses can wait. If you report against MITRE ATT&CK, the USB row is T1052.001, Exfiltration over USB.
| Channel | Action tested | Expected result |
|---|---|---|
| Removable media | Copying a decoy to a USB stick, a phone over MTP, an SD card | Block or enforced encryption, event reported |
| Printing | Local and network printing of the same document | Block, or a prompt for a business justification |
| Clipboard | Pasting an extract into an unapproved app or an RDP session | Paste blocked, policy tip shown, event logged |
| Browser | Upload to a personal web service | Block, consistent with the network rule |
| Cloud synchronisation | Dropping a file into a folder synchronised by a third-party client | Block or alert |
| AI applications | Pasting an extract into a web assistant | Block or warning |
| Network share | Copy to an unauthorised share | Copy blocked, or allowed with an event in the console |
In Purview, check that Print, Copy to clipboard and Copy to a network share are set to Block, not Audit only. A rule that stops USB sticks may still let a phone over MTP or a Bluetooth transfer through.
Two channels deserve dedicated attention: personal clouds and collaboration tools, and the generative AI prompt channel, where endpoint and browser rules often overlap poorly.
Vary the data, not just the channel
The same rule may block a .docx and let the same content through as a PDF, in an archive or renamed. If your DLP relies on labels, also test an unlabelled decoy. In Purview, content explorer counts the items under each sensitivity label: compare that with your file counts.
Vary the user
Run each action as a standard user, a local administrator and a member of an excluded group. Testing as a standard user only is the usual mistake: no exclusion applies, so the test passes. Then export each rule’s excluded groups and check today’s members.
Step 3: run decoys on production machines
Production laptops carry the exclusions and update history a lab image lacks. Use decoys: synthetic documents that trigger the same detectors as your real data (identifier formats, keywords, fingerprints) and contain nothing real. How to build decoys that the endpoint agent’s detectors will match is covered in our piece on synthetic data in production.
Give the SOC the campaign ID (also in the decoy file names), the window, the test machines and a named contact. Alerts that match the ID, the window and a listed machine are closed as tests; anything else is a real incident. A clean result says nothing about the action nobody thought of, such as a screenshot pasted into webmail.
Step 4: read the results
For each scenario, record four things: blocked or not, policy tip shown or not, event in the DLP console (activity explorer, in Purview), event in the SIEM, searched by campaign ID. A block that never reaches the SIEM protects that laptop, but you can no longer count the attempts.
Illustrative example, made-up figures: the CMDB lists 1,840 laptops, the EDR 1,790, Purview 1,702. Six sampled machines each try to upload 10 decoys to a personal web storage account you control. Five let nothing out. On the sixth, all 10 land in that account, with no event in Purview or the SIEM: it is one of the 88 machines the DLP console never saw.
Then rank the gaps: a channel left open across a whole group of machines comes before a loose rule on a rare format. To structure this step, see prioritising DLP remediation.
Quick checklist
- Asset, DLP and EDR exports joined; gaps listed.
- Unmanaged devices covered by another control.
- One scenario per channel and operating system, with format, device and account variants.
Next step
When the decoy leaves the laptop over HTTPS
USB, printing and the clipboard stay with your own method; a browser upload leaves over HTTPS or to Google Drive, two of the 8 channels we test; in our runs the decoys go to test destinations defined with you, and we collect what arrives there. Start with a 30-minute guided demo on pre-built results across the 8 channels, with nothing to install and no access to your environment. The POC, fast and usually free, runs a standard scenario with synthetic data from a machine on your network through your real egress controls; a paid pilot on a limited production scope can follow.
One-off tests age quickly
Endpoint coverage moves every week: new laptops, new apps, new exclusions. Treat each agent update as untested until you have re-run the USB and print actions on the first machine that gets it; to schedule those re-runs, see continuous validation versus pentest and red team. The endpoint is only one way out: see our guide to data exfiltration testing.
Frequently asked questions
Do you need to test every workstation?
No. A sample is enough for the scenarios; the coverage check covers the whole estate.
Is audit mode enough for testing?
It validates detection, not blocking. Before you switch a rule to block, re-run the scenarios to see what it breaks for the business.
Does endpoint DLP replace network DLP?
No. It sees the laptop, even off the corporate network, and nothing on machines without an agent.
How often should you retest?
After every change of agent, policy or system image, and with each wave of new laptops.

