Skip to content
Terms.

Amazon Bedrock FAQs changed AWS, Sep 9, 2026

AWS · Sep 9, 2026 · 24 added, 5 removed · found in an Internet Archive capture

Archived copyAll AWS terms
  1. Added: Amazon Bedrock Guardrails offers different safeguards to help you build safe, generative AI applications.
  2. Added: Multi modal content filters - Helps detect and filter harmful text and/or image content across multiple categories including hate, insults, sexual, violence, misconduct, and prompt attacks.
  3. Added: Denied topics - Define a set of topics that are undesirable in the context of your application. The filter will help block them if detected in user queries or model responses..
  4. Removed: Amazon Bedrock Guardrails offers six safeguards to help you build safe, generative AI applications. Below are the safeguards offered by Bedrock Guardrails.
  5. Removed: Multi modal content filters - Configure thresholds to help detect and filter harmful text and/or image content across multiple categories including hate, insults, sexual, violence, misconduct, and prompt attacks.
  6. Removed: Denied topics - Define a set of topics that are undesirable in the context of your application. The filter will help block them if detected in user queries or model responses.
  7. Added: Contextual grounding checks - Helps detect and filter hallucinations if the responses are not grounded (e.g., factually inaccurate or new information) in the source information and irrelevant to user's query or instruction.
  8. Added: Automated Reasoning checks - Helps detect factual inaccuracies in generated content, suggest corrections, and explain why responses are accurate by checking against a structured, mathematical representation of knowledge called an Automated Reasoning Policy.
  9. Added: For agentic applications, the InvokeGuardrailChecks API supports content filters, prompt attack detection, and sensitive information filters as standalone, resource-less checks. See "What is the InvokeGuardrailChecks API?" below.
  10. Added: The InvokeGuardrailChecks API is targeted towards agentic AI applications and lets you apply individual safeguards at any point in your application without creating guardrail resources. This API offers a unique approach to building safe agentic AI applications with the flexibility to invoke supported safeguards at any turn in the agentic loop, empowering you to take the required action in your application logic. There is no need to create any guardrail resources up front. The API returns numeric scores for each safeguard that enables you to define custom thresholds to block, bypass, or log results for auditing purposes based on your application requirements.
  11. Added: The InvokeGuardrailChecks API offers the following capabilities:
  12. Added: Resourceless: You don't need to create guardrails resources upfront. There's no CreateGuardrail step, no guardrail IDs to track, and no versions to manage. You specify which safety checks to run directly in each API request. This makes it straightforward to add, remove, or adjust checks as your workflows evolve.
  13. Added: Detect-only: The API does not block, mask, or rewrite content. It returns findings with numeric scores for each safeguard, and you decide what action to take based on your application requirements. This gives you full control to implement context-aware logic - blocking high-confidence threats, routing ambiguous findings to human review, or logging low-confidence results for audits.
  14. Added: Symmetric request-response: The safeguards you configure in your request are the same keys returned in the response. If you request contentFilter and sensitiveInformation, only those two will appear in the API results. This makes it easy to map findings back to the safeguards that produced them.
  15. Added: The InvokeGuardrailChecks API returns two types of numeric scores dependent on the safeguard that you apply.
  16. Added: Severity scores: With content filters and prompt attack detection, the API returns a severity score, which is a numeric value in the range [0, 1] that represents the strongest match for the safety check criteria. Specifically, the score matches content moderation or prompt attack detection. These scores are discrete with the possible values being 0, 0.2, 0.4, 0.6, 0.8, and 1.0. A score of 1 (maximum) is the strongest match, indicating the closest match to the definition of the safety check. A score of 0 (minimum) is the weakest match, indicating the content is benign. The severity score is a property of the content itself and not the certainty of the underlying model about its classification. It answers the question, "how strongly does this content match the criteria for the safety check?". It does not answer the question "how sure is the underlying model?" The score is agnostic in semantics but is specific to particular safeguards.
  17. Added: Confidence scores: With sensitive information filters, the API returns a confidence score that measures the model's certainty. This score indicates how certain the model is about its prediction. This is relevant to a PII detection on how confident the model is about the presence of PII entities as selected when the API is invoked. These scores are a series of continuous scores.
  18. Added: You can use either or both APIs depending on your use case. For example, use ApplyGuardrail for your customer-facing chat interface with automatic enforcement, and InvokeGuardrailChecks within your agent loop for granular, adaptive safety checks.
  19. Added: Applies targeted checks at specific points in agentic workflows.
  20. Added: Does not require creating, versioning, or managing guardrail resources.
  21. Added: Returns numeric scores (severity or confidence); you decide what action to take.
  22. Added: Best suited for agentic AI workflows with per-step safety requirements.
  23. Added: Provides uniform enforcement across your entire application.
  24. Added: Requires creating, versioning, and managing guardrail resources upfront.
  25. Added: Automatically blocks, masks, or bypasses content based on pre-configured thresholds; returns a pass/fail outcome.
  26. Added: Best suited for traditional request-response AI applications.
  27. Removed: Contextual grounding checks - Help detect and filter hallucinations if the responses are not grounded (e.g., factually inaccurate or new information) in the source information and irrelevant to user's query or instruction.
  28. Removed: Automated Reasoning checks - Help detect factual inaccuracies in generated content, suggest corrections, and explain why responses are accurate by checking against a structured, mathematical representation of knowledge called an Automated Reasoning Policy.
  29. Added: Note: Prompt attack is included as part of the content filter safeguard with the ApplyGuardrail API. With the InvokeGuardrailChecks API, prompt attack is a separate safeguard that you can apply while invoking the API. When using the prompt attack filter with the InvokeGuardrailChecks API, you can specify the individual type(s) of prompt attack giving you flexibility based on your application requirements.
About this change
Page
aws.amazon.com/bedrock/faqs/
Kind
Help article
Text hash
b5df2193047c to 794cd38212dc
Dated by
the first Internet Archive capture sampled that shows the new text; the change happened on or before this date

Terms changes by email

Mondays, only in weeks when a watched page changed.

Double opt-in. Unsubscribe any time.