subtopic

Input guardrails: personal data and harmful requests

Learn to check what goes INTO the model: mask personal data (names, phone numbers, Aadhaar/PAN, account numbers, emails) before sending text to an external API or storing logs; block or redirect off-topic and harmful requests with a moderation check or a small classifier prompt ('Is this question about company policy? yes/no'). Unmask only in the final answer if needed.