Machine Learning
Amazon Bedrock Guardrails
Amazon Bedrock Guardrails applies configurable safety policies to model inputs and outputs, including content, denied topics, sensitive information, and contextual grounding checks. It supplies a reusable defense layer across supported Bedrock interactions.
Key points
- Guardrail versions and identifiers make policy deployment traceable
- ApplyGuardrail can evaluate text outside a model invocation
- Intervention metadata should feed monitoring and evaluation
- In a Knowledge Bases runtime flow, retrieved references are outside the guardrail's input and generated-response checks
When to use it
- Block disallowed support content across several models
- Mask sensitive identifiers before returning generated text
Exam tips
- Choose Guardrails for model-content policy, not IAM, tenant filtering, or business authorization
- Tune against both adversarial and legitimate examples because aggressive filters can damage task completion
- Sanitize and govern sources separately; attaching a guardrail does not make retrieved documents trusted instructions