Defending LLM-based Multi-Agent Systems Against Cooperative Attacks with Sentence-Level Rectification | ArxivCSExplorer