PsySafe: A Comprehensive Framework for Psychological-based Attack, Defense, and Evaluation of Multi-agent System Safety
Zaibin Zhang (Shanghai Artificial Intelligence Laboratory), Yongting Zhang (Shanghai Artificial Intelligence Laboratory, University of Science and Technology of China), Lijun Li (Shanghai Artificial Intelligence Laboratory), Hongzhi Gao (Shanghai Artificial Intelligence Laboratory, University of Science and Technology of China), Lijun Wang (Dalian University of Technology), Huchuan Lu (Dalian University of Technology), Feng Zhao (University of Science and Technology of China), Yu Qiao (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Published Jan 22, 2024 ยท arXiv:2401.11880
cs.CLcs.AIcs.CRcs.MA
Abstract
The authors propose a comprehensive framework (PsySafe) grounded in agent psychology to identify how dark personality traits in agents can lead to risky behaviors, evaluate the safety of multi-agent systems from the psychological and behavioral perspectives, and devise effective strategies to mitigate these risks.
