The safest PII is the PII you never send. Most prompts that carry it don’t need it.
How it works
NIST SP 800-122 (April 2010) uses a broad definition: PII is “any information about an individual maintained by an agency, including (1) any information that can be used to distinguish or trace an individual’s identity, such as name, social security number, date and place of birth, mother’s maiden name, or biometric records; and (2) any other information that is linked or linkable to an individual, such as medical, educational, financial, and employment information.”
The GDPR’s term is personal data: “any information relating to an identified or identifiable natural person”, including “a name, an identification number, location data, an online identifier”. Saudi Arabia’s PDPL defines personal data in similar terms.
Redaction tools usually work in two steps. Microsoft Presidio’s analyzer handles “PII identification in text”, using patterns and named-entity recognition. Its anonymizer then applies an operator to each match:
| Operator | Presidio’s description |
|---|---|
| replace | “Replace the PII with desired value” |
| redact | “Remove the PII completely from text” |
| mask | “Replace the PII with a given character” |
| hash | “Hashes the PII text using salted hashing for security” |
| encrypt | “Encrypt the PII using a given key” |
A placeholder pattern for a support ticket:
Before: Ahmed Al-Harbi (+966 5x xxx xxxx) says order 4471 never arrived.
After: <PERSON_1> (<PHONE_1>) says order 4471 never arrived.
You keep the mapping on your side and put the real values back into the model’s answer.
Why it matters
- Every copy is a risk. NIST advises organisations to “minimize the use, collection, and retention of PII to what is strictly necessary”. Prompts and logs are copies.
- Detection misses things. Presidio says “there is no guarantee that Presidio will find all sensitive information”. Arabic names, Gulf phone formats and mixed-script text are where generic detectors often miss; test on your own data.
- Outputs can leak too. Run the same checks on responses. Guardrails can block or flag them.
Sending sensitive data to an LLM API covers redaction, placeholders and logging in detail, and usage logging with PII redacted covers your own logs.
On BroadNet Inference
BroadNet’s privacy statement says it does not retain prompts, audio, images, answers or transcripts, and does not use inference data to train, fine-tune or evaluate models. That limits what is stored after the request. It does not redact for you: the PII still reaches the model during processing, so minimise first. See zero data retention.
BroadNet also publishes an open-source redaction proxy, the LLM Sensitive Data Firewall (LSDF), under Apache-2.0; the sensitive data post describes it and BroadNet’s own evaluation figures. This page is not legal advice.