The checks around a model that constrain what it may be asked and what it may return — input filtering, output validation, refusal rules, and limits on which actions it can trigger.
A model in front of customers is a public surface. Without guardrails it will eventually quote a price it invented, repeat an injected instruction, or take an action nobody authorised.
We validate output against a schema, keep the model away from irreversible actions without a human step, and treat retrieved documents as data rather than instructions — which is what blocks prompt injection.