Guardrails Concepts — Safety and Guardrails

Defense Layers Around the Model

Amazon Bedrock Guardrails is a configurable, model-agnostic safety layer you can attach to any inference call via guardrailConfig or run standalone via ApplyGuardrail (no model invocation needed). It ships six safeguard types:

• Multimodal content filters (text + images) - hate, insults, sexual, violence, misconduct, and prompt-attack detection across both text prompts/outputs and uploaded or generated images, with per-category strength tuning.
• Denied topics - natural-language descriptions of subjects to block entirely before or after generation.
• Word filters - exact-match custom word/phrase blocklists (e.g. competitor names, profanity).
• Sensitive information filters (PII redaction) - managed PII entity detectors (SSN, email, credit card, health IDs) plus custom regex, with BLOCK or ANONYMIZE masking actions.
• Contextual grounding checks - score RAG outputs on both grounding (is the claim derived from retrieved Knowledge Base chunks?) and relevance (does it answer the user's prompt?).
• Automated Reasoning checks - translate natural-language compliance policies into formal mathematical logic, proving with up to 99% verification accuracy whether model outputs satisfy business and regulatory rules rather than relying on probabilistic guesses.

Attaching a Guardrail to Converse (boto3)

response = client.converse(
    modelId="us.anthropic.claude-sonnet-5-5",
    messages=[{"role": "user", "content": [{"text": user_prompt}]}],
    guardrailConfig={
        "guardrailIdentifier": "gd-ent-prod-01",
        "guardrailVersion": "1",
        "trace": "enabled",   # inspect exact policy verdicts in response["trace"]
    },
)

Guardrails ship in Classic and Standard safeguard tiers (Standard extends detection into code elements like comments and identifiers), support cross-Region inference distribution, and can be enforced cross-account. Do not rely on model behavior alone for safety-critical decisions.