Monitoring LLM-based Multi-Agent Systems Against Corruptions via Node Evaluation