Guardrails
Guardrails are technical or content-based safeguards that prevent an AI system from producing undesired, unsafe, or rule-violating output or actions. They range from simple content filters to structural rules governing which tools a model is even allowed to call.
Source: OpenAI — Guardrails