Guardrails (AI) in plain English.

Guardrails are the rules, filters, and checks wrapped around an AI system to keep its behavior within safe, intended limits — separate from the model's own built-in safety training. A guardrail might block a chatbot from discussing competitors, strip personal data from a response before it's shown, refuse to let an agent execute a destructive database command, or flag output that fails a content policy check before it reaches the user. Guardrails typically sit at three points: before the prompt reaches the model (input filtering), while the model is deciding what tool to use (action limits), and after the model responds (output filtering). Products like Guardrails AI, NeMo Guardrails, and Lakera build this as a standalone layer that works across any underlying model. The honest limitation: guardrails catch known failure patterns reliably but, like the models they wrap, are not foolproof against novel attacks.