Definition
Guardrails sit around a model rather than inside it. Input guardrails screen requests for abuse, sensitive data or prompt injection; output guardrails check answers for policy breaches, personal data or unsupported claims; action guardrails limit which tools an agent may call and with what values.
They range from simple rules and allow-lists to classifier models and dedicated frameworks such as NVIDIA's NeMo Guardrails. No single guardrail is complete, so they are layered, tested with red teaming and monitored in production.