Guardrails and safety filters
Also called: input/output filtering, safety layer, content moderation
Checks placed around a model — on what goes in and what comes out — to catch unsafe, off-policy, or malformed content before it reaches a user or an action. They wrap the model; they aren't part of it.
A model's own behaviour is probabilistic and steerable by its input, so you don't rely on it to police itself. Independent guardrails are how you enforce a hard line — blocked topics, injection patterns, malformed tool calls — that a prompt alone can't guarantee.
The system prompt. Telling the model to 'be safe' is guidance it may or may not follow; a guardrail is an external check that runs regardless of what the model decided.
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1U.S. National Institute of Standards and Technology · 2023-01-26
First-pass citations, limited to primary sources; a reviewer will broaden and verify these before this entry leaves draft.