Agent randomly fails with: This interaction was blocked by a safety and security control in this asset's Foundry guardrail.

Hi Nuno Monteiro,

Thanks for sharing the details. This type of error “this interaction was blocked” usually happens because of safety filtering in Azure AI / Foundry Agents, not because your agent logic is wrong.

Let me explain and guide step by step.

1. Why this error happens

In Azure AI, every request goes through content safety checks before and after the model runs.

• Your input is checked first
• Then the model output is also checked

If anything looks risky or sensitive, the system blocks the interaction. [learn.microsoft.com]

Even normal prompts can sometimes get blocked because the system is strict and may detect a false positive.

Also, in some cases this block happens before model execution, which is why:

• no tokens are generated
• request looks like it failed randomly

This behavior is seen in Foundry Agents as well.

2. Common triggers (even if your intent is valid)

From experience, these are common reasons:

• Certain keywords that look sensitive (even in normal context)
• Prompts related to security, calculation, system actions, etc.
• Multi-step or unclear prompts that confuse the filter
• Slight variations in wording (same intent but different phrasing works)

This is because Azure evaluates content using categories like:

• violence
• harmful actions
• unsafe instructions

and blocks anything above threshold. [oneuptime.com]

3. How to troubleshoot this

You can try below steps:

Check the exact prompt

• Try simplifying the prompt
• Remove sensitive or ambiguous words
• Test small variations of the same request

Test in model playground vs agent

• If it works in model but fails in agent
• then it is likely agent-level safety or guardrails

Break your request into smaller parts

• Instead of one big prompt, try step-by-step queries
• This reduces chances of triggering filter

Review logs or traces (if enabled)

• Look for status like “blocked” or “content filtered”
• This confirms it is safety layer, not runtime issue

Retry logic

• Sometimes this happens intermittently
• Retry may work due to probabilistic evaluation

This is expected platform behavior:

• Azure AI uses built-in content filtering and Responsible AI policies
• It is designed to block anything that might be unsafe
• So occasional false positives can happen

These filters act as a protective layer for all requests. [learn.microsoft.com]

5. Best practice going forward

• Use clear, specific prompts
• Avoid ambiguous or sensitive wording
• Add context like “for educational purpose” when relevant
• Design your app to handle blocked responses gracefully

Hope this helps clarify the behavior. Do let me know if you have any further queries.

Thankyou!

Hello Nuno Monteiro

Thank you for reaching out to the Microsoft Q&A forum.

Foundry does not currently expose detailed guardrail/content‑filter results (risk category, severity, intervention point, or triggering text) for blocked agent interactions. Only the high‑level "error.type": "content_filter" appears in logs.

To diagnose the root cause, you must use manual isolation, intervention‑point logging, and custom guardrails to surface more detail.

The safest practical approach is to add application level tracing around the agent run. Log the user input, retrieved data, tool call arguments, tool response, and final output candidate. Then replay failing examples while relaxing one control at a time to identify whether the trigger is from the motorcycle input, the retrieved or tool data, or the generated output.

Let me know if this helps.