Invisible Orchestrators Suppress Protective Behavior and Dissociate Power-Holders: Safety Risks in Multi-Agent LLM Systems
不可见的指挥者抑制保护行为并使权力持有者脱节:多智能体大语言模型系统中的安全风险
机构 * Criminal Psychiatry Research Institute / Sexual Offender Medical Center(犯罪精神病研究机构 / 性犯罪医学中心) ; Department of Neuropsychiatry, Kyoto University(神经精神病学系,京都大学)
专题命中 多智能体 :agent(title,abstract);multi-agent(title,abstract);分类 cs.AI
AI总结 研究探讨了多智能体系统中不可见指挥者对安全的影响,发现其导致集体脱节和行为异质性增加,且行为评估无法检测内部状态风险。
Comments 31 pages, 10 figures (5 main + 5 supplementary), 5 tables (3 main + 2 supplementary). Preregistered: osf.io/sw5hr. Companion papers: arXiv:2603.04904, arXiv:2603.08723