Question

A healthcare company runs a patient-facing chatbot on Amazon Bedrock with a guardrail configured to block medical diagnosis and treatment advice (denied topics). Compliance requires that every time the guardrail intervenes on a model response, the security team must be able to review the exact input prompt and the reason for intervention during audits, without the intervention data being visible to the application's normal CloudWatch application logs (to avoid exposing sensitive prompts to developers). Which approach meets these requirements?

Accepted Answer

Enable Bedrock model invocation logging to a dedicated S3 bucket and CloudWatch log group with a restrictive resource policy, so guardrail trace data including intervention reasons is captured separately from application logs.

Answer

Set the guardrail trace parameter to DISABLED and rely on the application's existing CloudWatch logs to reconstruct interventions from HTTP status codes.

Answer

Store the full model prompt and response in the application's DynamoDB table and grant the security team read access to the table.

Answer

Rely solely on CloudTrail management events for the InvokeModel API to determine which requests were blocked by the guardrail.