Compositional Unalignment: Auditing Safety in Large Vision-Language Modelswith Kai Shu, Li Xiong1 versionW_v9cspeeav1 · current