Source: Amazon Bedrock — Guardrails

Bedrock's configurable safety layer: content filters, denied topics, word and PII filters, contextual grounding and Automated Reasoning checks.

Contents

Summary

Catalogues the guardrail policy types and how a guardrail is versioned and attached to model calls or used standalone.

Key claims

  • Content filters (Hate, Insults, Sexual, Violence, Misconduct, Prompt Attack) with per-category strength, on text and images.
  • Standard tier extends detection into code elements (comments, identifiers, string literals) and improves grounding checks; Classic is the other tier.
  • Denied topics; word filters (case-insensitive whole-word/phrase); sensitive-information filters that block or mask PII, plus custom regex.
  • Contextual grounding checks flag responses not supported by the source or irrelevant to the query (for RAG).
  • Automated Reasoning checks validate responses against logical rules, flag unstated assumptions and suggest corrections.
  • Guardrails don’t evaluate reasoning content blocks.
  • Attach by guardrail ID + version on inference calls, or call ApplyGuardrail with no model; input tagging lets you evaluate only the user part of a prompt (SDK only).
  • Supports cross-Region guardrail inference and cross-account enforcement.

Feeds

Amazon Bedrock

Suggest changes →