ACIArena: A Unified Evaluation for Agent Cascading Injection

The researchers propose ACIArena, a unified framework for evaluating the robustness of multi-agent systems (MAS) against agent cascading injection (ACI) attacks. It covers three attack surfaces—external input, agent configuration, and inter-agent messages—and three attack goals—instruction hijacking, task disruption, and information leakage—and includes six widely used MAS implementations and 1,356 test cases. Benchmark results show that evaluating MAS robustness based on topology alone is insufficient; robust MAS require careful role design and controlled interaction patterns; defenses developed in simplified environments often struggle to transfer to real-world scenarios, and overly narrow defenses may even introduce new vulnerabilities.