Runtime checks that constrain LLM inputs and outputs to keep behavior safe and on-spec.
1 min readupdated 2026-06-21
/ quick answer
Guardrails are programmatic policies wrapped around an LLM: input filters (PII, prompt-injection detection), output validators (schema, toxicity, factuality), and fallback behaviors. They turn a probabilistic model into a system you can ship to production with predictable failure modes. Runtime checks that constrain LLM inputs and outputs to keep behavior safe and on-spec.
Runtime checks that constrain LLM inputs and outputs to keep behavior safe and on-spec. Guardrails are programmatic policies wrapped around an LLM: input filters (PII, prompt-injection detection), output validators (schema, toxicity, factuality), and fallback behaviors. They turn a probabilistic model into a system you can ship to production with predictable failure modes. In practice: Before a support agent sends a reply, a guardrail strips customer PII from logs, validates the reply matches a JSON schema, and blocks responses that violate the refund policy. This dictionary node is part of the Onexial knowledge graph and links to related concepts, workflows and tools below.
Definition
Guardrails are programmatic policies wrapped around an LLM: input filters (PII, prompt-injection detection), output validators (schema, toxicity, factuality), and fallback behaviors. They turn a probabilistic model into a system you can ship to production with predictable failure modes.
Example
Before a support agent sends a reply, a guardrail strips customer PII from logs, validates the reply matches a JSON schema, and blocks responses that violate the refund policy.
Guardrails are programmatic policies wrapped around an LLM: input filters (PII, prompt-injection detection), output validators (schema, toxicity, factuality), and fallback behaviors. They turn a probabilistic model into a system you can ship to production with predictable failure modes.
What is an example of Guardrails?
Before a support agent sends a reply, a guardrail strips customer PII from logs, validates the reply matches a JSON schema, and blocks responses that violate the refund policy.
Why does Guardrails matter for AI and automation?
Runtime checks that constrain LLM inputs and outputs to keep behavior safe and on-spec. It connects to the workflows, prompts and tool stacks linked on this page, so you can move from definition to execution without leaving Onexial.