A tutoring company launches a Bedrock chatbot for K-12 students needing blocked self-harm content, redacted PII, and grounded answers with citation enforcement. Parents require logged policy actions. Which generative AI control design enforces content filtering and compliance logging?

Bedrock Guardrails evaluate both inputs and outputs against denied topics such as self-harm, content filters, PII redaction, and contextual grounding checks that block responses not grounded in the supplied source passages, with intervention logging to CloudWatch for audit and tuning, providing deterministic policy independent of model wording. IAM boundaries, MFA, and SCPs control who can invoke Bedrock but cannot inspect or transform generated text for harmful content or privacy. CloudFront WAF with rate limiting mitigates web exploits and abuse at the edge but does not understand semantic topics, redact PII inside completions, or verify grounding. Prompt instructions alone are bypassable, non-auditable, and lack enforced logging of policy actions, failing compliance assurance. For the exam remember Guardrails sit on the inference path for filtering and logging, and in practice version guardrail configs, test adversarial prompts, and review blocked-request metrics regularly.