Output filtering
Also known as: output filter · response filtering
Output filtering inspects a model’s output before it’s shown or acted on, catching harmful, secret, or off-policy content.
- Blocks leaked secrets or PII.
- Catches toxic or disallowed content.
- A last line of defense before the response ships.
Layer it: filtering can't fix a bad model, but it catches the worst leaks. Pair it with a good system prompt and least privilege for real protection.