description [ACL 2026][LLM Safety][Multi-agent systems] This paper constructs the first unified evaluation framework for "Agent Cascading Injection (ACI)" attacks, ACIArena. It covers 6 mainstream ...