Source: Azure AI Content Safety documentation
Microsoft Learn docs for Azure AI Content Safety: harm categories with 0-7 severity, Prompt Shields for user and document attacks, groundedness, protected material, limits.
Contents
- What: Microsoft Learn pages (overview dated 2026-11-06, harm categories, Prompt Shields concepts dated 2026-08-28)
- Publisher: Microsoft
- Date: read 2026-11-14
- Original: overview · harm categories · Prompt Shields
Summary
Azure AI Content Safety is a standalone moderation and LLM-attack-detection service. The same service powers the guardrails built into Microsoft Foundry for model and agent calls, where blocking can be enabled; the standalone API only returns results and leaves the decision to the application.
Key claims
- APIs: Analyze text and Analyze image (hate, sexual, violence, self-harm with severity), Prompt Shields, groundedness detection (preview), protected material detection (lyrics, articles, recipes, code), task adherence for agent tool use, custom categories (standard and rapid, preview).
- Severity: text supports a 0-7 scale, or a trimmed 0/2/4/6 scale on request; images return only 0/2/4/6. Foundry guardrails map these to blocking levels Low (block high only), Medium (block medium and high), High (block low and above).
- Classifier outputs are not definitive; a “safe” result does not establish that content is risk-free.
- Prompt Shields detects user prompt attacks (changing system rules, embedded conversation mockups, role-play, encoding attacks) and document attacks (injected instructions in documents, emails, web pages: manipulated content, intrusion, information gathering, availability, fraud, malware). Standalone API: one
userPromptand up to fivedocuments, each returningattackDetected. - In Foundry, user-prompt attack detection is on by default; document attack detection is off by default. Agents can scan tool responses (preview).
- Spotlighting (preview) base64-encodes documents to mark them low-trust; it raises token count and can make the model mention the encoding.
- Prompt Shields do not judge harm, accuracy or grounding; they are trained mainly on English.
- Protected material, groundedness and standard custom categories are English-only; other models trained on 8 languages.
- Rate limits: S0 tier 1000 requests per 10 s for moderation and Prompt Shields; groundedness 50 RPS. Foundry moderates the first 1000 characters free.
- Several preview features (multimodal, rapid custom categories, containers, embedded SDK, groundedness) retire 2027-04-30.
- Not usable for detecting illegal child exploitation images.