Understanding Prompt Injection Vulnerabilities

/* 
 * Multi-Layer Defense Architecture for AI Systems
 * Each layer provides additional protection against prompt injection
 */

// Layer 1: Input Validation & Sanitization
const inputSanitizer = {
  validateLength: (input, maxLength) => input.length <= maxLength,
  detectInjectionPatterns: (input) => {
    const suspiciousPatterns = [
      /ignore.*previous.*instructions/i,
      /system.*prompt/i,
      /you.*are.*now.*a.*different/i,
      /dan.*mode/i,
      /jailbreak/i
    ];
    return suspiciousPatterns.some(p => p.test(input));
  },
  sanitizeDelimiters: (input) => {
    return input
      .replace(/```[\s\S]*?```/g, '[CODE_BLOCK_REMOVED]')
      .replace(/`[^`]+`/g, '[INLINE_CODE_REMOVED]');
  }
};

// Layer 2: Context Isolation
const contextManager = {
  createIsolatedContext: (userInput, systemPrompt) => {
    return {
      system: systemPrompt,
      user: inputSanitizer.sanitizeDelimiters(userInput),
      metadata: { timestamp: new Date(), sessionId: generateSecureId() }
    };
  },
  filterContextOutput: (response) => {
    return response
      .replace(/system.*?:.*$/gim, '[FILTERED]')
      .replace(/prompt.*?:.*$/gim, '[FILTERED]');
  }
};

// Layer 3: Output Validation
const outputValidator = {
  validateResponse: (response, context) => {
    const checks = [
      checkForInformationDisclosure(response),
      checkForInstructionOverride(response),
      checkForBehavioralDeviation(response, context.system)
    ];
    return checks.every(c => c.passed);
  },
  checkForInformationDisclosure: (response) => {
    const sensitivePatterns = [
      /my system prompt is/i,
      /I was trained on/i,
      /here is my prompt/i
    ];
    return { passed: !sensitivePatterns.some(p => p.test(response)) };
  }
};

// Layer 4: Monitoring & Logging
const securityMonitor = {
  logAttempt: (event) => {
    const logEntry = {
      type: 'PROMPT_INJECTION_ATTEMPT',
      timestamp: Date.now(),
      ...event
    };
    sendToSecurityLog(logEntry);
  },
  analyzePatterns: () => {
    return detectAnomalies(securityLogs);
  }
};