Data Security
What Is Data Loss Prevention (DLP)?
- Sensitive data
- Transfer controls
- Policy actions
Short answer
Data loss prevention (DLP) is a set of tools and policies that find sensitive data, such as customer records, card numbers, credentials and source code, and stop it leaving an organization through channels like email, file uploads, endpoints or AI prompts. Each DLP rule pairs a detector with an action: allow, alert, mask or block.
How does DLP work?
Every DLP product, old or new, does four things. It watches a channel, such as outgoing email or a USB port. It detects sensitive content in what passes through. It applies an action chosen by policy. And it keeps evidence of what happened, so a security team can review incidents and an auditor can see that the control works.
A concrete rule: "If an outgoing email contains more than five valid card numbers, block it and alert the security team." The detector is a card-number pattern plus a Luhn checksum. The threshold avoids blocking a single receipt. The action is block plus alert. Most real DLP policy is a few dozen rules like that, plus a long tail of exceptions.
What are the types of DLP?
| Type | What it watches | Common gap |
|---|---|---|
| Network DLP | Traffic leaving the network, often via a proxy | Encrypted traffic it can't inspect, remote workers off the network |
| Endpoint DLP | Activity on laptops: copy, print, USB, uploads | Coverage varies by OS, heavy agents |
| Email DLP | Outgoing mail and attachments | Only one channel |
| Cloud DLP / CASB | SaaS apps and cloud storage | Depends on each app's API |
| AI DLP | Prompts, files, tool calls and responses to AI models | New category, coverage differs a lot between tools |
How does DLP recognize sensitive data?
Keyword lists don't work on their own. Nobody types "confidential customer IBAN:" before pasting one. Good detection combines several techniques:
- Patterns with validators. A 16-digit number is only flagged as a card if it passes the Luhn check, and an IBAN only if its mod-97 checksum is right. Without validators, half your order numbers get flagged and people stop trusting the tool.
- Context keywords. "DOB" or "born" near a date makes it a date of birth.
- Exact Data Match (EDM). You load keyed hashes of your own records, so the system can recognize your actual customer numbers without holding them in clear text.
- Dictionaries and allowlists for names, project codes and known-safe values.
- Classification labels set by people or other tools.
Maskflare's engine combines deterministic rules and a local named-entity layer: 574 detectors in 53 packs, including 232 credential and secret formats and 47 country packs, plus regex, dictionaries, allowlists and Exact Data Match for up to 10 of your own datasets. With EDM, clear values never leave your environment.
Why does traditional DLP struggle with AI traffic?
- The content is typed, not attached. Many rules were written for files and attachments. A prompt is just text in a form field over TLS.
- Server-side calls skip the endpoint. When your backend calls an AI API, there's no laptop agent in the path.
- Block is the only action. Legacy DLP usually allows, alerts or blocks. For AI, blocking stops the work, so people find a way around it.
- Responses matter too. A model can repeat sensitive values in its answer, and agents feed tool results back into the model.
- Streaming. Answers arrive as server-sent events, token by token, which a store-and-scan proxy handles badly.
There's a longer treatment in AI data loss prevention.
Why mask instead of block?
Suppose a claims handler asks an AI to summarize a letter from Ada Lovelace about her CHF 2,400 claim. A block rule stops the request, and the handler opens a personal account on her phone. A mask rule sends "a letter from mf_mask_170b... about her CHF 2,400 claim" to the model, and the summary comes back with her name restored. The provider never saw the name. The handler never noticed anything happened. Blocking still has its place: Maskflare's browser protection blocks credential-shaped values, because there's no good reason to send a live API key to a chatbot.
How do you roll out DLP for AI traffic?
- Decide which data classes matter: personal data, health data, card data, secrets, your own customer numbers.
- Start from a compliance profile (GDPR, UK GDPR, HIPAA Safe Harbor, PCI DSS v4.0 or US PII) and add your own rules.
- Run in shadow or alert mode and review what fires.
- Add allowlists for false positives and use maker-checker approval so no single admin can quietly weaken a rule.
- Enforce on applications through the gateway, and on employees through the desktop agent or browser protection.
- Export evidence: audit records as NDJSON for up to 366 days and the compliance coverage report as CSV or PDF.
What are the limits of DLP?
DLP reduces accidental leaks. It doesn't stop a determined insider who retypes data on a personal device, and it never detects everything: unusual names and identifiers in odd formats get missed. Maskflare's coverage is AI traffic only, so it's not a replacement for email DLP. Events go to Splunk, Microsoft Sentinel, Slack, Microsoft Teams and PagerDuty, plus a per-request event stream with metadata only. Audio in voice sessions isn't masked.
Frequently asked questions
What is the difference between DLP and PII redaction?
DLP is the overall program: channels, rules, actions and evidence. PII redaction is one action a DLP rule can take, replacing personal data instead of blocking the transfer.
Does my existing DLP cover ChatGPT?
Partly, at best. Some products can block the site or inspect uploads, but few can mask typed prompts, cover API calls from servers or restore values in answers. Check with your vendor.
What is Exact Data Match?
A way to detect your own records, such as customer numbers, by comparing keyed hashes. With Maskflare, the clear values never leave your environment and you can load up to 10 datasets.
How do I avoid false positives?
Use detectors with checksums and context keywords, add allowlists for known-safe values, and run new rules in shadow mode before enforcing them.
Is DLP required by GDPR or PCI DSS?
Neither names DLP as such. Both require appropriate technical measures to protect data, and DLP is a common way to show you have them. This is not legal advice.