An artificial intelligence safeguard is a control placed around data, models, tools, outputs or decisions. Examples include access limits, validation, human approval, monitoring, rate limits, replay tests, rollback and restrictions on how a system may act.
A production safeguard should make a known unsafe path difficult or impossible rather than depend only on instructions. Each control needs a defined failure mode, evidence that it operates and an owner who can respond when it is triggered or bypassed.
