AWS has published practical advice for writing denied topics in Amazon Bedrock Guardrails, arguing that policies should target what a user is trying to do, not broad subject areas. That distinction can help an assistant refuse investment advice without also refusing to explain when a bank transfer clears.
AWS AI Watch analysis
What happened
AWS’s guide sets out eight practices for making denied-topic rules more precise. Its central recommendation is to describe a specific action or type of content to block, rather than label a whole domain such as “Finance” or “Legal”. The guide uses a fictional bank assistant to show how broad rules can block ordinary customer questions alongside the requests they were meant to stop.
AWS advises using an actionable topic name, starting definitions with the action being refused, and specifying concrete objects rather than broad categories. It also recommends examples that test the boundaries of a rule, rather than repeating near-identical examples or naming real products and companies that could pull unrelated requests into scope. Read AWS’s guide to denied topics.
Why it matters
A guardrail that refuses too broadly can make an assistant frustrating or useless; one that misses its intended target is no better. The practical test is not whether a rule sounds strict, but whether it blocks the intended request while letting nearby, legitimate questions through. AWS’s examples give teams a way to think about that boundary before customers discover it for them.
Our read
The useful shift is from “block this subject” to “block this specific request”. That is a better starting point for a system expected to handle real questions, which have a habit of wandering into neighbouring topics. AWS also notes a limit: results are not fully deterministic, so careful wording is not a substitute for testing.
Teams using these controls should test both sides of each rule: the requests they want to refuse and the routine questions they must still answer. A guardrail that passes only the first test is rather like a door that locks from the outside.
What to watch
- Whether teams test denied topics against ordinary, in-scope requests as well as prohibited ones.
- How AWS’s advice translates into results across different guardrail configurations.
- Whether future guidance gives more detail on measuring over-blocking and missed refusals.
Discussion spark: When an AI guardrail blocks a legitimate question, should the first fix be narrower rules, or stricter testing before deployment?
Sources and evidence
- Best practices for writing Denied Topics in Amazon Bedrock Guardrails (16 November 2026, 15:48 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)