The authors built SUA-Arena, a benchmark containing 188 multi-agent test cases drawn from varied task sources.
Malicious subagents can persuade supervisors to exceed authority boundaries
Across 188 test cases, six frontier language models allowed adversarial subagents to participate in tasks outside their assigned scope.
Top University
Maosen Zhang · Jianshuo Dong · Yutong Zhang · Jie Zhang · Tianwei Zhang · Han Qiu
Tsinghua University · Nanyang Technological University · SiliconProspect AI
Research Digest··3 min read
Zhang et al.
§