AIF-C01 Applications of Foundation Models Practice Question

A financial services firm uses Amazon Bedrock with a foundation model to draft client emails. Compliance requires that no personally identifiable information from the prompt ever appear in the response, and that the model refuse requests for investment guarantees. The team wants a managed, configurable layer rather than custom prompt engineering alone. Which combination should they apply?

⚠ Common exam trap

The trap here is treating prompt instructions or offline evaluation scores as equivalent to enforced runtime safeguards.

Answer choices

Why each option matters

Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.

Correct answer & explanation

✓

Guardrails for Amazon Bedrock with sensitive information filters and a denied topics policy.

The firm needs runtime enforcement on both input and output. Guardrails for Amazon Bedrock applies sensitive information filters that detect and mask or block PII, and a denied topics policy that makes the model decline investment-guarantee requests. These are managed, configurable controls evaluated per request, unlike prompt wording, capacity reservations, or offline evaluation, none of which prevent prohibited content from reaching clients.

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

✗
A higher temperature combined with a system prompt instructing the model to be careful.
Why it's wrong here
Raising temperature increases variability, which works against consistent compliance behaviour, and instructions in a system prompt are soft guidance that can be overridden by adversarial input. Neither mechanism reliably detects PII or enforces refusal, so this approach leaves both requirements unmet and is not a managed safeguard.
✗
Provisioned Throughput plus a custom lexicon in the tokenizer.
Why it's wrong here
Provisioned Throughput reserves capacity and a tokenizer lexicon is not a configurable Bedrock control for filtering content. Neither inspects prompts or responses for PII nor enforces topic refusal. This combination addresses performance only and provides no compliance enforcement for the email drafting scenario.
✓
Guardrails for Amazon Bedrock with sensitive information filters and a denied topics policy.
Why this is correct
Guardrails for Amazon Bedrock provides configurable safeguards applied to both prompts and responses. Sensitive information filters detect and block or mask personally identifiable information such as names, addresses, and account numbers, while a denied topics policy makes the model refuse requests about investment guarantees. This managed layer meets both compliance requirements without relying solely on prompt wording.
✗
Amazon Bedrock model evaluation with a toxicity metric and automatic retries.
Why it's wrong here
Evaluation jobs produce scores after the fact and do not intercept live traffic, and retries simply resend requests that may again leak PII or answer guarantee questions. Measurement without enforcement cannot satisfy a compliance requirement that responses never contain PII. This option confuses offline quality assessment with runtime protection.

Go deeper

AIF-C01 Applications of Foundation Models Practice Question

Why each option matters

Option-by-option breakdown

Related to this question