Agent Action Firewall Documentation

Key Concepts

Actions

Actions are requests from AI agents to perform operations like HTTP calls, database queries, or external API interactions. Each action is validated against security policies, evaluated for risk level, routed for approval if needed, and logged in the audit trail.

Policies

OPA (Open Policy Agent) policies written in Rego determine how actions are handled:

• Allow: Action proceeds immediately
• Deny: Action is blocked with reason
• Require Approval: Action waits for human approval

Approvals

High-risk actions can require human approval before execution. Approvals can be managed through the dashboard, Slack, Discord, Teams, or integrated with Jira/ServiceNow for ticket-based workflows.

Audit Trail

Every action, decision, and approval is logged with cryptographic hash chaining for tamper evidence. Proof packs can be exported for compliance and forensic analysis.