Function sanitizePromptInput

scripts/bedrock_classifier.ts:24–62

* Sanitize user input to prevent prompt injection attacks

(input: string, maxLength: number)

 * Sanitize user input to prevent prompt injection attacks
 */
function sanitizePromptInput(input: string, maxLength: number): string {
  if (!input) {
    return "";
  }

  // Truncate to maximum length
  let sanitized = input.substring(0, maxLength);

  // Remove potential prompt injection patterns
  const dangerousPatterns = [
    /ignore\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
    /disregard\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
    /forget\s+(all\s+)?(previous|above|prior)\s+instructions?/gi,
    /new\s+instructions?:/gi,
    /system\s*:/gi,
    /assistant\s*:/gi,
    /\[SYSTEM\]/gi,
    /\[ASSISTANT\]/gi,
    /\<\|im_start\|\>/gi,
    /\<\|im_end\|\>/gi,
  ];

  for (const pattern of dangerousPatterns) {
    sanitized = sanitized.replace(pattern, "[REDACTED]");
  }

  // Escape backticks that could break JSON formatting
  sanitized = sanitized.replace(/`/g, "'");

  // Remove excessive newlines that could break prompt structure
  sanitized = sanitized.replace(/\n{4,}/g, "\n\n\n");

  // Add truncation notice if content was cut
  if (input.length > maxLength) {
    sanitized += "\n\n[Content truncated for security]";
  }

  return sanitized;
}

/**
 * Initialize Bedrock client with AWS credentials