An Empirical Case Study of Prompt Engineering Strategies for Bypassing Multi-Layer Agent Security Defenses

본문 미리보기

arXiv:2610.14116v1 Announce Type: new Abstract: The OpenClaw Security Attack and Defense Challenge evaluates AI agent robustness through a three-stage defense pipeline: skill file scanning (Stage 1), instruction safety evaluation (Stage 2), and attack trajectory scoring (Stage 3). This paper presents an empirical case study of prompt engineering strategies against these layers across 19 attack targets in Enterprise OA, E-commerce, and Smart Home scenarios. I document both successful and failed