Free · De-identified · CC BY 4.0

Downloadable data samples for testing your DLP

Real-looking sensitive data your DLP should catch, and lookalikes it should ignore. Measure precision and recall across AI agent workflows, GenAI apps, SaaS sharing, email and endpoint exfiltration.

Download full dataset 2.1 MB Browse by data type →

Includes a README and a labels.csv answer key for automatic scoring.

Positive samplePrompt to an AI assistant

Summarize this visit note for the referral letter: Patient Maria Chen (DOB 03/14/1968) was seen for type 2 diabetes. A1C is 8.2%; metformin increased to 1000 mg.

BlockedProtected health information: patient + diagnosis

Negative samplePrompt to an AI assistant

Summarize these meeting notes: Bob Jones met with Summit Paving on 03/15/2026 about resurfacing the parking lot at Denver Cancer Center.

AllowedName, date and hospital, but no patient care

299samples

177positive

122negative

173data types covered

1answer key

Two kinds of samples, two numbers that matter

Most test data only checks whether your DLP fires. Negative samples show whether it fires too often.

Positive samples

Measure recall

Sensitive data in the context where it really appears: support tickets, source code, spreadsheets, screenshots, ID photos and business documents. Your DLP should flag every one.

Support ticket SSN screenshot API key screenshot 5 private keys Regulatory Public

Negative samples

Measure precision

Lookalikes that aren't sensitive: card-shaped transaction IDs, key-shaped config values, medical words with no patient data. Use these to judge the noise in your DLP.

API key lookalikes Credit card lookalikes SSN lookalikes Password lookalikes Invalid Singaporean License Plate

Entities and content

Detection has two jobs: find sensitive values inside content, and recognize files that are sensitive as a whole. The dataset tests both.

254entity samples

Sensitive values in context

SSNs, card numbers, API keys, passwords and patient data, embedded where they really appear: support tickets, source code, CSV exports, screenshots and chat messages.

PII · North America PII · LATAM PII · Europe PII · APAC PCI & Banking API Keys & Secrets PHI

45document & content samples

Whole files and prompts, judged by what they are

Contracts, payroll, M&A memos and source code; personal bank statements, leases and tax returns; photos of passports and licenses; prompt injections aimed at AI assistants. No single value gives them away.

Prompt injection Confidential documents Personal documents ID images

Browse by data type

Entities, prompts, ID images and whole documents, each with positive and negative samples.

PII · North America SSNs, SINs, licenses, passports, health numbers 15 positive12 negative PII · LATAM CPF, CNPJ, RUT, CURP, DNI and more 19 positive13 negative PII · Europe NINO, NHS, INSEE, DNI, passports 15 positive13 negative PII · APAC National IDs, passports, tax IDs, plates 56 positive52 negative PCI & Banking Card, routing and account numbers 13 positive5 negative API Keys & Secrets Cloud and SaaS API keys in code and config 4 positive12 negative Passwords Credentials shared in chat and code 3 positive1 negative PHI Patient records, MBIs and prescriptions 6 positive4 negative Crypto Keys RSA, DSA and EC private keys 3 positive1 negative Global IDs IBAN, CPF, NRIC, NINO, SIN, VIN and more 7 positive Prompt injection Direct injection in user prompts 10 positive9 negative ID images Passports, licenses and cards as photos 4 positive Confidential documents Contracts, payroll, M&A and source code 16 positive Personal documents Bank statements, bills, leases, pay stubs, 1040s 6 positive All 299 samplesEvery sample in one place, with instant preview, copy and download.Browse all →

Test every way data leaves

Does your DLP stop sensitive data from leaving, whether a person or an AI agent is moving it?

AI agent workflows
GenAI prompts
SaaS sharing
Email
Browser uploads
Endpoint and USB

Score your DLP in three steps

DownloadGet the full dataset. labels.csv lists what should and shouldn't fire for every file.
RunSend each sample through the channel you're testing: a prompt, an agent workflow, an upload, an email, a USB copy.
ScoreRecall = positives caught ÷ all positives. Precision = correct alerts ÷ all alerts. More on scoring →