Azure AI Content Safety: Prompt Shields vs blocklist vs Task Adherence vs custom categories vs multimodal
Verdict: Hidden instructions in documents need Prompt Shields. Fixed banned terms need a blocklist. Misaligned tool calls need Task Adherence. Emerging brand patterns need custom categories. Image plus caption needs the multimodal API.
| Criterion | Prompt Shields | Custom blocklist | Task Adherence | Custom categories rapid | Multimodal analyze API |
|---|---|---|---|---|---|
| Detects or does | Instructions hidden in user prompts or third-party documents | Blocks a fixed set of terms regardless of severity score | Flags agent tool calls or responses misaligned with the user task | Scans text and images for emerging brand-specific patterns | Scores an image together with its caption in one call |
| Choose when | A RAG or agent pipeline ingests third-party or OCR'd document text | Reject specific competitor or internal code names | An agent invokes tools with inputs not matching intent | Brand-specific prohibited symbols | Combined image-plus-caption moderation |
Rules
- Use Prompt Shields indirect and document detection for instructions hidden in documents or OCR'd image text used as grounding.
- Use a custom blocklist as an input or output filter to block fixed terms regardless of severity scoring.
- Use Task Adherence to catch tool invocations that do not match the user's stated request.
- Use custom categories (rapid) to define and scan text and images for emerging brand-specific harmful patterns.
- Use the multimodal analyze API to score an image and its accompanying text together in a single call.
Traps
- Groundedness detection checks output against sources; it does not detect hidden instructions or block fixed terms.
- Protected material detection matches known copyrighted content, not malicious instructions or custom brand names.