PII Redaction
Also: data minimization · PII masking · sensitive-data scrubbing
Detect and remove personal data before it reaches the model, logs, or downstream systems that don't need it.
Why CCAR-P tests this
The exam tests whether you minimize exposure of personal data across prompts, logs, and outputs rather than trusting the model to keep it safe.
PII redaction is detecting personally identifiable information — names, emails, government IDs, card numbers, health data — and removing or masking it at the boundaries where it isn't needed. It is a concrete expression of data minimization: the less sensitive data flows through a component, the less can leak from it.
Redaction applies at multiple points and connects to guardrail placement: scrub PII on input before it enters the prompt or is sent to a third-party model, scrub it from logs and traces so debugging data isn't a breach waiting to happen, and screen output so the model can't emit PII to an unauthorized recipient. Where the model genuinely needs an identifier, prefer tokenization (a reversible placeholder) so the raw value never travels further than required.
Redaction is a named control in the register — with an owner and evidence (detection coverage, test results) — and it should fail closed: if the detector errors on a field, treat the data as sensitive rather than passing it through in the clear.
Exam trap
"We log full prompts and responses for debugging" is the trap — unredacted logs quietly become the largest PII exposure surface in the system.
Commonly confused with
Related concepts
Independent, unofficial study material from Claude Cert Prep. Not affiliated with Anthropic.