Multi-Agents Assisted Autonomous Breach and Attack Simulation Framework in Power Information System

Abstract

The growing prevalence of advanced persistent threats targeting power information system has made proactive security assessment and early vulnerability identification increasingly important to mitigate the potential impact of cyber-attacks. Nevertheless, accurately reproducing realistic red-team operations remains challenging due to the inherent complexity and long-term nature of APT activities. To address this issue, this paper introduces an end-to-end red teaming framework built upon cooperative multi-agent collaboration, enabling full life-cycle attack simulation in power information systems. Through coordinated interactions among multiple intelligent agents, the proposed system enhances the intelligence and coverage of attack simulations, constructs effective attack chains, and facilitates the discovery of deeply hidden security weaknesses. Experimental evaluations conducted on representative power information system scenarios demonstrate that the proposed framework substantially expands the diversity of executable attack actions while achieving a significantly higher exploitation success rate than existing approaches. In particular, it attains approximately fivefold improvement in attack action coverage, thereby enabling more comprehensive exploration of potential attack paths. Such enhanced capability makes the framework well-suited for realistically emulating advanced persistent threat scenarios and effectively identifying deeply hidden vulnerabilities within complex power information networks.