AIP-C01 practice question: Protecting PII and sensitive data (Guardrails sensitive info filters, Macie, Comprehend)

AI Safety, Security, and Governance A free question from our original AIP-C01 practice exam. Try it, then open the answer to see why each option is right or wrong.

An insurance company runs a claims assistant for its adjusters. Adjusters paste claim notes into the chat, and the assistant drafts replies using a model on Amazon Bedrock.

Generated replies sometimes repeat full Social Security numbers, card numbers and internal claim reference codes copied from the notes. Adjusters need the rest of each reply unchanged, and the company does not want an entire reply refused because of one value.

The control must mask Social Security and card numbers in every reply, and mask internal claim codes that follow a known format. The team accepts a small configuration change to its Converse calls but does not want to write or run any scanning code.

Which solution meets these requirements with the LEAST custom code?

Choose one answer.

A. Add a system prompt that tells the model never to repeat numbers that look like personal identifiers, and have a supervisor check a random sample of replies each day for any values that leaked.
B. Store every reply in an Amazon S3 bucket, run Amazon Macie on the bucket each night, and delete any object in which Macie finds personal data, so that stored records do not keep the identifiers.
C. Attach a guardrail to the Converse calls with sensitive information filters set to ANONYMIZE for Social Security and card numbers and a custom regex for claim codes.
D. Call Amazon Comprehend to detect personal data in each reply, and return a generic error message in place of any reply that contains a detected identifier, so that no personal data is shown.

Show the answer and explanation

Answer: C

A. Wrong: A prompt instruction is not an enforced control, and the model can still repeat identifiers in its reply. Daily sampling finds leaks after adjusters have already seen them, and it does nothing to mask the values.
B. Wrong: Macie finds sensitive data at rest in S3, so it works on stored replies after the fact. Adjusters would still see the raw values on screen, and deleting whole objects removes replies the team needs.
C. Right: The sensitive information filters replace matched values in the reply with placeholders, and the custom regex covers the internal claim format. The rest of each reply is returned unchanged, and the only change is the guardrail setting on the Converse calls. Learn more: Sensitive information filters in Amazon Bedrock Guardrails (opens in a new tab)
D. Wrong: Comprehend detects personal data well, but returning an error hides the whole reply, which the company does not want. It also adds a second service call and custom code in every application.

Explanation

Guardrail sensitive information filters can anonymize matched values in the reply, and a custom regex covers internal identifiers, so the rest of each answer is still returned. Prompt instructions and after-the-fact scans leave values visible, and replacing whole replies with errors removes useful content.

AIP-C01 practice question: Protecting PII and sensitive data (Guardrails sensitive info filters, Macie, Comprehend)

Explanation

Study

Keep practicing