Glossary · Evaluation & safety
Guardrails
System controls that constrain inputs, tool use, outputs, permissions, and escalation. They can include schemas, policy checks, classifiers, allowlists, sandboxing, approvals, and post-action verification.
Why it matters
No single filter covers all failure modes, so controls should be layered according to risk.
Common confusion
Guardrails reduce risk; they do not prove that an AI system is safe.
Related terms
Browse the learning paths to see this term in context — every lesson is free to read.