Reproduce “ACIArena: Toward Unified Evaluation for Agent Cascading Injection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hengyu An, Minxi Li, Jinghuai Zhang, Naen Xu, Chunyi Zhou, Changjiang Li, Xiaogang Xu, Tianyu Du and 1 more
ACIArena evaluates multi-agent security using 1,356 cases across three attack surfaces and objectives. It pairs benign tasks with attacks and measures task utility alongside attack success. In a controlled CORBA test, attack success varied by agent profile, while topology trends were inconsistent. ACI-SENTINEL reduced hijacking ASR to 0.00% in Code-domain MetaGPT, but benign utility fell; findings assume one malicious agent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.