AI guardrail
Also known as: guardrail · AI safety control · runtime control
A guardrail is a control that stops an AI system from producing or executing something unwanted, applied at the moment of execution.
Two families are often conflated. Content guardrails filter what the model produces — prompt injection, data leakage, prohibited speech. Action guardrails decide whether an operation may happen at all.
They also differ in where they live. Some install inside the agent's execution; others stay outside and return an opinion the infrastructure applies. The first requires controlling the platform, the second does not.
Neither replaces the other: a content filter has no view on whether a cross-border transfer is lawful, and an authority decision does not stop a model from saying something foolish.
What it means for a small business
A small business buying off-the-shelf agents usually has no control over execution. The realistic guardrail is then external: a call before the sensitive action.
Related terms
So what actually applies to you?
A definition tells you what a term means, not what your organization must do. The assessment answers the second question — free, no credit card.
Updated September 1, 2026 · Educational definition; not legal advice.