AI Explained
PII redaction for LLM apps: mask it in, restore it out
Keep names, emails and card numbers out of your prompts. Detect personal data, swap it for placeholders, keep the mapping, and restore the real values in the answer.
On this page
A support ticket hits your LLM app. It contains a customer's name and their card number. Now what?
Where personal data leaks
Once a prompt leaves your system, that data can end up in logs, in a vendor's system, or back in the model's output. OWASP lists this risk as LLM02:2025, Sensitive Information Disclosure, and it explicitly includes personally identifiable information (PII).
The fix is simple to state: redact before the prompt leaves your system, and restore after the answer comes back.
Step 1: detect
Tools like Presidio scan text for entities such as PERSON, EMAIL_ADDRESS, PHONE_NUMBER, CREDIT_CARD and IBAN_CODE. They mix methods:
- Named entity recognition for names,
- patterns and checksums for numbers (a credit card number has to pass a checksum),
- context words nearby, like "card" or "phone".
Step 2: transform
Once found, each value can be replaced with a label, redacted, masked (4111 •••• •••• 1111), hashed, or encrypted.
| Operator | Reversible? |
|---|---|
| Replace / redact / mask | No (unless you keep a mapping) |
| Hash | No, one way |
| Encrypt | Yes, with the key |
The placeholder trick
For LLMs, placeholders work especially well:
- Sarah becomes
<PERSON_1>, the card becomes<CREDIT_CARD_1>. - You keep the mapping on your side. The model only ever sees the labels.
- When the answer says "Thank
<PERSON_1>", you swap the real name back in.
A bonus: the model provider never receives the real values, so there's less to leak. Keep labels consistent (same person, same label) so the model can still follow who did what. Presidio ships a sample that does exactly this, with the caveat that its simple version isn't thread-safe.
No detector is perfect
Presidio's own docs warn: "there is no guarantee that Presidio will find all sensitive information." So add layers: limit what goes into prompts in the first place, and scan what comes out.
Know the law
Under GDPR, pseudonymised data is still personal data: Recital 26 says it "should be considered to be information on an identifiable natural person." Only truly anonymous data falls outside the regulation. Placeholders with a mapping you can reverse are pseudonymous, not anonymous.
Where to redact
- Prompts
- Logs and traces
- Anything that goes into training data (OWASP: "prevent user data from entering the training model")
Mask it on the way in, restore it on the way out.