In one sentence

Guardrails

Checks and rules around an AI that try to stop unwanted inputs, outputs, or actions, while still needing testing because they can miss things.

From the offline library