Definition
A guardrail is a control intended to keep an AI system's inputs, outputs, or actions within an acceptable range. It might be a rule, classifier, policy check, permission boundary, approval step, rate limit, sandbox, or rollback mechanism.
The word is broad enough to hide important differences. An advisory guardrail warns. A detective guardrail records or flags. A preventive guardrail blocks. A containment control limits damage after failure, and a recovery control helps reverse it. A useful system description says which job the control performs and what can bypass it.
Guardrails are strongest when they are enforced outside the model at the point of effect. A prompt that asks the model not to exceed a spending limit is guidance. A payment service that rejects transactions above the limit is an enforcement boundary.
Distinguish it from nearby terms
Policy states what should be allowed. A guardrail implements, checks, or supports that policy. A security boundary prevents unauthorized effects even when another component misbehaves. Not every guardrail qualifies as one.
Check your understanding
An output filter flags prohibited text but still sends it to a downstream tool. It is a detector, not a blocker. Calling it a guardrail without naming that limitation overstates the protection.