Classify sensitive data before it enters a prompt, avoid common data leakage mistakes, and recognize AI-specific threats like prompt injection and deepfake phishing.