Our centrally managed AWS Bedrock guardrails are blocking team requests, especially through the PROMPT_ATTACK policy. Even after lowering that policy's strength to LOW, requests are still being rejected. CloudTrail previously showed useful information about why a request was blocked, but that detail seems to have disappeared recently, leaving us unable to determine which guardrail is being triggered. Has anyone found a reliable way to diagnose these blocks?
2 Answers
Are you using streaming, particularly Converse Stream? There appears to be a Bedrock guardrails issue where streaming can cause false positives for malicious-intent and prompt-attack detection. We opened an AWS support case, and they reproduced the problem. They said they were working on a change to the chunking strategy used by guardrails during streaming, with a fix expected soon.
That sounds like a combination of false positives and reduced diagnostic detail. Guardrails are useful when they catch genuine attacks, but ordinary workflows become difficult to troubleshoot when the blocking policy is not reported. I’d check whether the issue is limited to streaming requests and ask AWS to confirm whether there has been a recent change to CloudTrail or Bedrock guardrail event logging.
Thanks, I’ll raise this with our AWS contact. Have you also noticed that CloudTrail stopped showing useful guardrail-trigger details recently? We could see them last week, but now the reason and specific policy are missing.