Zero Trust for Multi-Agent AI Systems

Abstract

Traditional Zero Trust asks systems to avoid granting implicit trust because of network location or asset ownership. Autonomous AI introduces additional boundaries: one agent can delegate to another, select a tool, call an API, act through a service, and affect a digital or physical resource. In that environment, authenticating the first participant does not establish that every downstream action is legitimate. This article develops a public, high-level SGAEIA model for applying Zero Trust to multi-agent systems. It separates identity from authority, capability from permission, delegation from trust propagation, and model reasoning from policy enforcement. It then derives practical architectural questions for continuous authorization, evidence generation, revocation, and operation at the edge.

This page provides the author-written abstract and bibliographic metadata for discovery. The Zenodo record is the persistent source for citation, files, version, and license information.