Azure Content Safety

Azure Content Safety provides enterprise-grade content moderation as a managed Azure AI service — covering text (up to 10,000 chars) and images with severity scores (0–6) across four harm categories: hate/fairness, sexual, violence, and self-harm. For agents building LLM-based applications: the Prompt Shields API detects jailbreak attempts and indirect prompt injection in user inputs. The Groundedness Detection API validates whether an AI response is grounded in provided context (for RAG pipelines). Custom blocklists allow organization-specific policy enforcement. Strong fit for enterprises already on Azure. Confidence is docs-derived.

REST API at {endpoint}/contentsafety/text:analyze?api-version=2024-09-01 (text) and /contentsafety/image:analyze (image). POST /text:analyze body: {text: '...', outputType: 'FourSeverityLevels'}. Response: {categoriesAnalysis: [{category: 'Hate', severity: 2}, {category: 'Sexual', severity: 0}, ...]}. POST /image:analyze: {image: {content: '{base64}'}} or {image: {url: '...'}}. Prompt Shields: POST /text:shieldPrompt. Groundedness: POST /text:detectGroundedness. Custom blocklist: POST /text/blocklists/{name}/blocklistItems. All endpoints versioned via api-version query parameter.

Two auth paths: (1) API key: Ocp-Apim-Subscription-Key header. Keys from Azure Portal → Content Safety resource → Keys and Endpoint. (2) Azure AD: Bearer token via DefaultAzureCredential. HTTPS enforced. Regional endpoint per Azure resource (e.g., eastus.api.cognitive.microsoft.com). Managed Identity supported for Azure-hosted services. Key rotation: Azure Portal or CLI. Per-resource keys (not per-user).

Standard HTTP status codes with Azure AI error envelopes ({error: {code, message}}). 401 for invalid API key, 429 for rate limit (Retry-After header), 400 for content exceeding size limits (text: 10,000 chars max). Severity 0 = safe; severity 6 = highest severity. Configurable blocklistMatchResults for custom blocklist detections. Azure Content Safety SLA: 99.9% availability (standard Azure AI SLA). Region-specific; use the same region as your resource. Latency: ~200–500ms for text; ~1–2s for images.

learn.microsoft.com/azure/ai-services/content-safety provides comprehensive documentation — quickstarts, REST API reference, concept guides for each feature (text analysis, image analysis, prompt shields, groundedness). Getting started: create Azure subscription, provision Content Safety resource, grab API key, first call in under 10 minutes. Official SDKs: Python (azure-ai-contentsafety), .NET, Java, JavaScript. Azure CLI for resource management. Active Microsoft Learn community.