PII Redaction
Stripping or masking personally identifiable information from data before it reaches a model, a log, or a third party.
PII redaction detects and removes personal data — names, emails, phone numbers, government IDs — from text before it flows into a prompt, a training set, or an application log. It’s a required control anywhere an LLM pipeline touches user data, both because prompts sent to third-party model APIs are a real exfiltration surface and because logs are often retained (and read by more people) far longer than anyone expects.
Worked example: detecting and masking personally identifiable information (names, emails, SSNs, card numbers) before it is logged, sent to a third-party model, or stored — usually via regex for structured formats plus an NER model for names/addresses. Gotcha: regex catches well-formed cases (email, card via Luhn check) but misses context-dependent PII (a name in free text), while an over-eager redactor destroys useful data; redaction after the log write is too late — do it before data leaves your trust boundary.