Place Guardrails within an application safety boundary
Separate content checks from document authorization and business-action controls.
Draw four different control points
Input checks handle inappropriate requests and sensitive input. Retrieval authorization controls which documents a caller may use. Output checks assess generated content. Tool authorization controls business actions. These checks solve different problems; a content filter does not grant record access or approve a payment.
Prepare an allowed-use policy, prohibited content examples, sensitive-data rules and the application's data flow. Map each rule to the component that enforces it. Decide what users see when a request is blocked, partly masked or sent for human review.
Configure only supported checks
Bedrock Guardrails offers configurable content, denied-topic, word and sensitive-information filters, plus grounding and other checks where supported. The standalone ApplyGuardrail API can evaluate supported text independently of model invocation. Direct integration depends on model, API, Region and endpoint.
For example, the current GPT-6.1 Sol card excludes integrated Guardrails with Responses, despite broader API documentation. Do not assume that adding headers enables protection on every path. Select a documented supported path or an explicit application-level check, then test it.
Evaluate false positives and missed cases
Use allowed, disallowed, ambiguous and multilingual samples. Include a prompt injection in retrieved content, sensitive data split across messages and a normal business request resembling a prohibited topic. Record the policy version, expected decision and actual decision.
With streaming output, decide whether partial content can reach the user before assessment completes. Verify the real buffering and error behavior rather than assuming token streaming is already safe.
Deliver the policy-to-control map, configuration references, sample results and escalation path. Keep permission and action tests alongside filter tests; no configuration establishes absolute safety or legal compliance.
Official references
Documentation checked: 17 October 2026. The instructions are reference guidance; no cloud API was called for this article.