AI guardrails are runtime constraints applied to model inputs and outputs — blocking, redacting, rewriting or escalating content that violates policy — operating on the request and response rather than on the underlying data.
What are AI guardrails?
Guardrails are necessary but frequently over-credited. They inspect what is going into and coming out of a model, which makes them effective against unsafe phrasing, obvious prompt injection and recognisable sensitive strings. They are far weaker against exposure created upstream, because a guardrail cannot detect that a conclusion was inferred from context the requester was never entitled to. Guardrails should be treated as the last layer of context governance, not a substitute for governing retrieval, memory and tools.