Securing Cross-Enterprise AI Agents

Abstract

As enterprises adopt autonomous AI agents, they need a way to preserve human accountability while allowing agents to act across organizational boundaries. This research presents a secure reference architecture for cross-enterprise AI agent delegation that separates who initiated an action from who executed it. The architecture propagates human and agent identity, validates delegated authority, enforces policy at each trust boundary, and records auditable evidence of agent actions. Using emerging and established standards such as RFC 8693, ID-JAG, W3C Verifiable Credentials, OPA, OpenTelemetry, and AGNTCY, the work outlines a practical Phase 1 proof of concept for delegated authorization, identity propagation, and cross-domain trust. The result is a foundation for more secure, governable, and interoperable enterprise AI agent ecosystems.