Five policy types carry the everyday work, one line of defence each: what is harmful, what is off limits, what is banned by name, what is private, and what the source does not support. A sixth, Automated Reasoning checks, validates an answer against formal logic extracted from a written policy. It returns findings rather than blocking, and it rarely comes up at this level.
The material lists the features of responsible AI as “bias, fairness, inclusivity, robustness, safety, veracity”. A guardrail covers two of those directly. Content and word filters serve safety, catching toxicity before it reaches anyone. The contextual grounding check serves veracity, blocking an answer the source does not support. The rest go to other tools, which is the sorting a scenario usually turns on. Grounding is the one candidates forget, so pair this card with the wider set of moves against a confident wrong answer. Configuring the policies for real goes well past what this level asks for.
Flash card
Amazon Bedrock Guardrails: a safety layer applied to prompts and responses independently of the model, so the same policy applies when the model changes.
- Content filters cover five harm categories, hate, insults, sexual content, violence and misconduct, plus a sixth for prompt attacks on the input. Each is set to none, low, medium or high, separately for prompts and responses. Raising the strength filters more harmful content and more innocent text with it.
- Denied topics are written in natural language: a noun-phrase name, a definition of up to 200 characters, and up to five optional sample phrases, 30 per guardrail. A retail assistant is kept off legal advice without anyone listing trigger words.
- Word filters match an exact list you supply, up to 10,000 entries of three words each, plus a managed profanity list AWS maintains. Competitor names are matched literally, not classified.
- Sensitive information filters detect personally identifiable information [PII]. Each type, an email address, a phone number, a card number, is set to block the request or mask the value with its type name, such as {EMAIL}. A custom regular expression covers an internal account number.
- The contextual grounding check scores a response twice: grounding, how far the answer follows from the source passed with the request, and relevance, how far it answers the query. Thresholds run from 0 to 0.99; a score below either blocks the response. It runs on output only.
Pick it when
Pick it when a generative feature faces people outside the team. One policy attaches to model inference, a knowledge base, an agent and a flow node. Guardrails are versioned, so a draft can be tested before a version goes live. The ApplyGuardrail API applies the same policies to any text, with no model invoked. Reach for it when a requirement is a rule the assistant must not break: a crafted prompt can override a system prompt.
It's the wrong answer when
It measures nothing. A guardrail enforces at runtime, one request at a time, so it cannot tell you whether a model treats demographic groups differently. Bias and fairness come from offline evaluation, subgroup analysis and human review. It is also not access control, which is AWS Identity and Access Management [IAM], nor network isolation, which is a virtual private cloud [VPC] endpoint.