Azure Content Safety within Azure Foundry - Not always blocking the same prompt, why?

Question

Azure Content Safety within Azure Foundry - Not always blocking the same prompt, why?

Isidro TTW 5

Hi,

Were using a separate subscription for AI, where we have a Llama-4 model (Azure Foundry), with some Azure Content Safety (Guardrails) adjustments, to "Content harms", including a blocklist (with "Exact match"'s).

According to our tests (prompts), even after adding the blocklist, several times the result/output passes (not blocked), and the tests include words that are in the blocklist.

Could you help explain why the results are not always blocked?

Thanks

0 comments

2 answers

Your answer

Answer 1

Chirag Kolte 0

Hi there,

Based on your description, I'd check the following:

Verify that the blocklist is associated with the same Azure AI Foundry project/deployment you're testing.
Confirm that the blocklist is configured to inspect the correct content (prompt, response, or both, depending on your scenario).
Since you're using Exact match, ensure your test output contains the exact blocked term. Even minor differences (such as punctuation, spacing, or word variations) may not match.
If you recently updated the blocklist, wait a few minutes and test again to allow the configuration to propagate.

If the exact blocked term is still returned after these checks, it would be helpful to know whether you're testing through Azure AI Foundry Playground or via the REST API/SDK, as that can help determine whether the issue is configuration-related or a service issue.

I hope this may help a lot

Isidro TTW 5 Reputation points

2026-07-23T13:57:16.2166667+00:00

Hi @Chirag Kolte ,

Thanks for your reply.

All the adjustments, including the blocklist, are linked to the correct model/project/deployment and were made some time ago.

The blocklist is currently inspecting both user input and output, will do some extra test following your suggestion.

The test were mainly via API, but via playground we also get ambiguous results.

Thanks.
Isidro TTW 5 Reputation points

2026-07-24T07:20:32.3433333+00:00

Hi again,

Did a few extra tests, and whenever were using "simple prompts", that include words in the blocklist, they are blocked, even if the words aren't a exact match (eg. 1st capital letter), but when we add Instructions (Playground) to the same prompts, we get answers (not blocked).

Our use case, sends alongside the user prompt, extra info, simulated it with the Instructions (Playground), so it seams we have found a way to bypass Content Safety, but we don't want to...Thanks

Answer 2

Deleted

This answer has been deleted due to a violation of our Code of Conduct. The answer was manually reported or identified through automated detection before action was taken. Please refer to our Code of Conduct for more information.

Comments have been turned off. Learn more