Guardrails are a set of safety and governance mechanisms designed to control the inputs and outputs of Large Language Models (LLMs) to ensure they remain secure, ethical, and aligned with organizational policies.

Within an AI Gateway, Guardrails can be static (using regex for pattern matching) or LLM-based (using semantic analysis) to identify and mitigate risks such as hallucinations, toxic content, and data leaks.

They enable real-time monitoring and intervention by blocking, signaling, or masking sensitive information, including Personally Identifiable Information (PII), thereby providing a critical layer of compliance and interaction security for Generative AI applications.