GLOSSARY · AI SECURITY
Guardrails
Guardrails are the technical controls wrapped around an AI model to constrain what it will accept and produce, separate from the model's own training.
Typical guardrails include input and output filters, topic and action allowlists, PII redaction, and policy checks on tool calls. They are defense in depth: model refusal training alone is not reliable.