AI-103 Practice Question 3: Custom Blocklist Terms
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 1 · question 3 of 10
A pharmaceutical company runs an internal assistant on a Foundry model deployment. Legal has supplied a list of about 60 unreleased drug codenames and an internal merger project name that must never appear in a model response, even though none of them is hateful, sexual, violent, or self-harm content; they are simply confidential strings. The default content filtering categories never flag them, because nothing about them is harmful.
What is the appropriate way to make the service flag content containing those specific terms?
Show answer and explanation
Answer:
The default AI classifiers cover general harm categories and are sufficient for most moderation needs, but they cannot know what is confidential inside one company. Blocklists let you add custom terms to the AI classifiers, so you can screen for specific terms or phrases that you want to flag in your content. A list of codenames is precisely the "specific to your use case" case blocklists exist for.
- B is wrong because severity thresholds only change how aggressively the *existing* harm categories fire. A drug codename carries no hate, sexual, violence, or self-harm signal at any severity, so lowering the threshold flags unrelated content instead. - C is wrong because Prompt Shields target adversarial instructions, user prompt attacks and indirect attacks embedded in grounding data. The problem here is disclosure of known strings, not an injected instruction. - D is wrong because protected material detection matches against known third-party copyrighted text and code, not against an arbitrary list of terms you supply.
Reference: Use blocklists for text moderation