Guardrail actions

Guardrail actions are what a guardrail does when a policy matches. The common ones are block, rewrite, redact, hold for approval, terminate the session, flag and alert, and each policy gets its own.

The actions

ActionWhat it doesWhen it fits
BlockStops the message, reply or tool callClear and serious violations
RewriteReplaces the reply with one that compliesA reply that is useful once fixed
RedactMasks a span and passes the restPersonal data and secrets
Hold for approvalPauses a tool call until a person confirmsAgent actions that cannot be undone
TerminateEnds the sessionAn attack in progress
FlagPasses it and marks it for reviewNew policies and borderline cases
AlertNotifies a person or a systemEvents someone must see now

How to choose

Match the action to the cost of a mistake. A new policy can start on flag and move to block once its false positives are low. A tool call has to be stopped or held before it runs. Holding it puts a human in the loop.

With White Circle

A policy in Control can block, rewrite, redact, hold a tool call for confirmation, terminate the session, flag, or send an alert.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL