What is Guardrails?

Input/output validation layers around an LLM that detect and block harmful content, prompt injection attempts, PII leakage, or off-topic responses. Typically a pipeline: input filter -> LLM -> output filter. Can be rule-based (regex, keyword lists) or model-based (classifier that scores safety).

What people say

Safety filters for AI

Why it's called that