In plain words
Guardrails are controls around an AI system that check or restrict what it receives, produces, or does. They can include input checks, output filters, tool restrictions, and approval steps.
A closer look
Different controls address different risks: an output check may detect sensitive information, while a tool permission can prevent an unauthorized operation. Some checks use rules; others use models.
Guardrails need testing on realistic requests and failure cases. Overly broad checks can block useful work, while weak checks can miss harmful behavior.
In practice
A support assistant can draft a refund, but the application requires approval before processing one above a set amount.
A useful distinction
A prompt asking the model to behave safely is not an enforceable permission boundary. Guardrails reduce risk but do not guarantee correctness or safety.