Guardrails & Responsible AI
Content filtering, PII detection, prompt injection protection, and responsible AI practices.
Authoritative Sources
Key guidance documents from authoritative organizations. Click to view the original source.
Bedrock Guardrails provides configurable safety and privacy safeguards across models. It supports content filters, denied topics, word filters, sensitive information filters, and grounding checks to help reduce harmful or irrelevant outputs in applications.
Configuration Examples(2)
Kwatra & Kaushik (Packt, 2024) address three categories of ethical risk in Bedrock workloads: veracity (hallucinations mitigated via prompt engineering, RAG, and temperature tuning), intellectual property (training-data provenance, content filtering, and watermarking), and safety/toxicity (guardrails, data curation, and responsible AI policies). The authors recommend combining Bedrock Guardrails with organizational responsible-AI governance frameworks.
Customer Configuration Responsibilities
Configuration tasks the customer owns in the shared responsibility model. Use the verification commands below to validate settings.
5. Guardrails Configuration
Apply content safety and privacy safeguards to model interactions.
Content filters
Configure input/output strength for content categories.
Denied topics
Define disallowed topics with definitions and examples.
Sensitive information filters
Configure PII detection and custom regex patterns.
Word filters
Block specific terms and managed word lists.
Contextual grounding
Set grounding and relevance thresholds for RAG outputs.
Blocked messaging
Define custom responses for blocked inputs or outputs.
Guardrail enforcement
Configure account or organization enforcement of guardrails.
Cross-account guardrail sharing
Use resource-based policies to share guardrails across accounts.
6. Application Security (Prompt Injection Protection)
Secure your application code and prompt handling around Bedrock.
Input validation
Validate and sanitize user input before invoking Bedrock.
Secure coding practices
Use parameterized queries and avoid unsafe string concatenation.
Security testing
Perform SAST, DAST, and penetration testing for AI workloads.
Agent pre-processing prompts
Tune pre-processing templates to classify and sanitize inputs.
System prompts and scope
Define what agents can and cannot do in system prompts.
SDK/library updates
Keep Bedrock SDKs and dependencies current with patches.
Verification Commands
Commands and queries for testing and verifying security configurations.
Guardrails
3 commandsaws bedrock list-guardrailsaws bedrock get-guardrail --guardrail-identifier GUARDRAIL_IDaws bedrock list-enforced-guardrails-configuration