Microsoft Foundry Generative AI Chat App Practice Question 3: Custom Guardrail Severity
Microsoft Applied Skills: Develop a Generative AI Chat App Using the Microsoft Foundry SDK · Test 1 · question 3 of 10
A Foundry project hosts two model deployments. One powers a public help bot; the other summarizes workplace incident reports, and its legitimate descriptions of injuries are being flagged by the Violence control of the default Microsoft.DefaultV2 guardrail. The public bot must keep its current protection unchanged. What should you do?
Show answer and explanation
Answer:
A guardrail in Microsoft Foundry is a named collection of controls, and each control defines a risk, the intervention points to scan (user input and output for models), and the response action. An individual guardrail "can be applied to one or many models," so the fix for a per-deployment false-positive problem is to author a separate guardrail with an adjusted severity threshold for the affected risk and assign only that deployment to it. The documented troubleshooting path for legitimate content being blocked is exactly to "review severity level settings for affected risk category" and "test with different severity levels to find appropriate threshold."
- A is wrong because Microsoft.DefaultV2 is the default guardrail assigned to models. Loosening it would also loosen the public help bot, which the stem forbids. - B is wrong because a model deployment cannot simply opt out of content safety. For Models sold by Azure, "only customers who have been approved for modified Guardrails can turn them off," through the Limited Access Review; it is not a self-service setting. - C is wrong because evaluators are an offline testing and monitoring feature that scores outputs against test data. They run as part of an evaluation, not inline on live requests, so they cannot change what a guardrail blocks.
Reference: Guardrails and controls overview in Microsoft Foundry