AI agent security glossary

Action and enforcement terms

Browser agent security governs what an AI agent may do through a browser. Action mediation evaluates a proposed action before execution. Pre-execution enforcement applies the policy outcome before browser state changes.

• Allow: permit the proposed action
• Ask: hold the action until a person decides
• Block: stop the action before execution
• Redact: remove sensitive data from the action or context
• Substitute: provide an approved value at the point of use

Credential and approval terms

Credential protection keeps secrets outside model context. Human-in-the-loop approval pauses a sensitive action until a reviewer explicitly allows or blocks it.

Evidence and context terms

An action audit trail connects the proposed action, target, session context, matched policy, decision, outcome, and reviewer history.