AI Glossary

Guardrails

What does “Guardrails” mean?

Guardrails: Controls placed around or inside an AI system to keep its behaviour within acceptable limits. They can include safety training, system instructions, input and output filters, topic restrictions, and rules limiting what tools an agent may use. Layering several guardrails is common because each one alone can fail.

Guardrails visual guide: Controls placed around or inside an AI system to keep its behaviour within acceptable limits. They can include safety training, system instructions, input and output filters, topic restrictions, and rules limiting what tools an agent may use. Layering several guardrails is common because each one alone can fail.
Visual guide Open full size (new tab)

What is an example of “Guardrails”?

A customer-service bot with guardrails might decline medical questions and block any reply containing a customer's full card number.

Keep learning

See also

Safety classifierTermJailbreakTermAgent guardrailsTermResponsible AITerm

Same topic

Frontier safety frameworkTermLeaderboardTermFairnessTermLLM-as-a-judgeTermExplainabilityTerm