MAStrike: Shapley-Guided Collusive Red-Teaming on Multi-Agent Systems

The article introduces MAStrike, a novel framework for collusive red-teaming in hierarchical multi-agent systems (MAS), addressing the limitations of existing heuristic approaches. It employs agent-level Shapley value analysis to quantify each agent's contribution to system robustness, enabling the identification of vulnerable agent coalitions and generating coordinated adversarial manipulations. The framework is validated through extensive experiments across various MAS architectures, revealing critical vulnerabilities and coordination patterns, which are essential for enhancing the security of high-stakes workflows in domains like finance and software engineering.