Content guardrail intermittently blocking a Quick flow

Q&A
, , ,
1

Hi team,

I have a daily Quick Flows flow that pulls open Jira cards, prioritizes them, and posts an executive briefing to Slack. It runs unchanged, but the “Prompt safety” guardrail keeps blocking it with “The input violates our content policy filters.”

Seems I’m not alone: Critical issue.

Why this looks like a non-deterministic false positive:

  • 57 runs, no changes: 38 Completed, 19 Failed (~33%). Same input, different outcome. A truly “bad” card would fail 100% of the time, not 1 in 3.
  • The flagged text is ordinary business content (e.g., a card title with an investment amount, standard KPI/OKR tasks). Nothing harmful.

Why I can’t fix it myself:

  • Thresholds aren’t user-configurable. The console only lets me add blocked words (tighten), never relax the built-in filter. No whitelist/exemption in the UI.
  • Renaming the flagged card is just a band-aid: tomorrow another card could get randomly flagged.

Questions:

  1. Any way to tune or get an exemption for the built-in guardrail?
  2. Is ~33% failure on identical input a known issue?
  3. Recommended pattern for legitimate business flows hitting these false positives?
2

Hi @gabriel.maia,

On tuning or getting an exemption, there isn’t a way to do that today. The blocked-words list only lets you tighten filtering, and there’s no whitelist or exemption for the built-in guardrail, so this isn’t something you can configure your way around.

For whether the ~33% failure on identical input is a known issue, I can’t confirm that specific intermittent pattern is formally tracked. What I can point to is the Limitations of apps in Quick documentation, which says, “Prompts occasionally trigger security guardrails as false positives.” So the phenomenon of legitimate content being flagged is acknowledged, and the non-determinism you’re seeing on unchanged input lines up with that.

And for recommended pattern question, I do want to be transparent that the options are limited since none of this is user-configurable. The one thing worth trying is seeing whether a specific recurring term or format is more likely to trip it, because if you can isolate a trigger you can sometimes reword.

Since this isn’t user-fixable and it’s blocking a legitimate business flow, this is one of the cases where raising it with AWS Support is worthwhile, I’d recommend opeining a support ticket (Case management - AWS Support). The guardrail behavior sits on the backend, and since this is a public-facing community, we do not have access to anyone’s account or haev the means to fix it directly on the backend.

Hope this helps, and if you do manage to isolate a particular trigger, sharing it here would help others running into the same thing!