content-safety-implementation

Content Safety Implementation

Implementation patterns for Azure Content Safety API integration, multi-layer defense pipelines, and operational safety controls for AI-facing applications.

Last validated: April 2026 (Prompt Shields GA, Groundedness Detection, Custom Categories)

---

Azure Content Safety API

import ContentSafetyClient from '@azure-rest/ai-content-safety';

interface ContentSafetyConfig {
  endpoint: string;       // from Key Vault
  apiKey: string;         // from Key Vault
  thresholds: {
    hate: 'low' | 'medium' | 'high';
    sexual: 'low' | 'medium' | 'high';
    selfHarm: 'low' | 'medium' | 'high';
    violence: 'low' | 'medium' | 'high';
  };
}

Threshold Selection Guide

Category	Low (strict)	Medium	High (permissive)
Hate	Consumer apps, children	General audiences	Historical fiction, education
Sexual	Most applications	Dating/health apps	Medical/clinical
Self-Harm	Default — always strict	Crisis support apps	Clinical research
Violence	Most applications	News, crime fiction	Medical, forensic

Usage Pattern

async function analyzeContent(text: string, config: ContentSafetyConfig): Promise<SafetyResult> {
  const client = ContentSafetyClient(config.endpoint, { key: config.apiKey });
  const result = await client.path('/text:analyze').post({
    body: { text, categories: ['Hate', 'Sexual', 'SelfHarm', 'Violence'] },
  });

  return {
    safe: result.body.categoriesAnalysis.every(
      c => c.severity <= SEVERITY_MAP[config.thresholds[c.category.toLowerCase()]]
    ),
    categories: result.body.categoriesAnalysis,
  };
}

---

Prompt Shields

Prompt Shields detect two attack types that bypass basic content filters:

User Prompt Attacks

Direct injection in user messages — role override, encoding tricks, conversation mockup, role-play manipulation.

const shieldResult = await client.path("/text:shieldPrompt").post({
  body: {
    userPrompt: userMessage,
    documents: retrievedDocs  // RAG context
  }
});

// Check for attacks
const { userPromptAnalysis, documentsAnalysis } = shieldResult.body;
if (userPromptAnalysis.attackDetected) {
  // Block: user prompt contains injection attempt
}