Execution Identity in Multi-Agent Systems: A Zero-Trust Runtime Governance Architecture (智能體系統的執行期身分:零信任運行時治理架構)

As Multi-Agent Systems proliferate across enterpriseand personal assistant applications, AI Agents are being widely deployed for retrieval and reasoning tasks in highly confidential environments. However, current mainstream frameworks universally rely on fragile “prompt-based soft boundaries” to control permissions, leading to a critical “Agentic Trust Paradox. ” Because Large Language Models inherently suffer from Prompt Injection and other incurable vulnerabilities, once an agent is compromised, it can easily forge identities and trigger unauthorized access with catastrophic security consequences. To contain such threats and establish accountability mechanisms, this paper proposes the DROS (Deterministic Runtime Operating System) architecture. Unlike traditional application-level Workload Identity, this research introduces a finer-grained “Execution Identity” model. Through dynamically bound, By-Execution digital certificates issued by the Agent Identity Authority (AIA), combinedwith the GuardVM zero-trust gateway, DROS successfully intercepted all attacks in our prototype evaluation. The architectureachieves cryptographic Non-Repudiation traceability with minimal latency overhead, establishing an infrastructure standard for the emerging secure “AgentWeb” era.-----------------------------------------------------------------------------------------------------------------------------------隨著多智能體系統(Multi-Agent Systems)