WTF is Guardrails?
Checks around an AI system that block, filter or fix unsafe or invalid inputs and outputs.
Also called: safety rails, output validation
TL;DR
Guardrails are the seatbelts: input checks (is this a jailbreak? does it contain personal data?), output checks (is this valid JSON? off-topic? leaking secrets?), and action limits (this agent may never delete production data).
How it works
Implemented as rule-based validators, classifier models, schema validation, or a second LLM acting as a checker, running before the model, after it, or around individual tool calls. Effective agent guardrails are mostly about permissions and blast radius (sandboxing, allow-lists, budgets, approvals) rather than trying to filter every bad string.
Origin
- Coined / introduced by
- Industry term
- When
- 2023
- Story
- Became common vocabulary in 2023 as companies shipped LLM features into production.