A company deploying a customer-facing generative AI assistant on Amazon Bedrock must prevent it from returning disallowed content or leaking internal data through crafted prompts. Which control addresses these model-layer risks?
- AAn AWS WAF web ACL with managed rule groups inspecting the model's inference API endpoint
- BAn IAM policy limiting which principals can invoke the model
- CAmazon Macie scanning prompts for sensitive data at rest
- DAmazon Bedrock Guardrails enforcing content and topic policies on prompts and responses
Bedrock Guardrails enforce content and topic policies on prompts and responses, mapping to LLM risk guidance. WAF, an IAM invoke policy, and Macie do not address model-layer content risk.
Why each option is right or wrong.
- D (correct). Bedrock Guardrails enforce policies on prompts and responses, addressing model-layer risk.
- A. WAF matches HTTP request patterns and does not evaluate prompts or responses against content and topic policies.
- B. An IAM invoke policy controls who can call the model, not the content of prompts or responses.
- C. Macie classifies stored data, not live model safety.
Reference: Bedrock Guardrails; SCS-C03 Domain 3.