An artificial intelligence guardrail enforces boundaries around model or agent behavior. Guardrails can validate inputs and outputs, restrict tools and data, require approvals, enforce policy or stop actions that do not satisfy defined conditions.
Guardrails reduce risk but are not complete guarantees. They need tests, monitoring and updates because models, tools and failure modes change, and a control that blocks one pattern may miss a different unsafe route.
