📄 中文摘要
随着大型语言模型（LLM）代理从基于文本的顾问演变为使用工具的自主行动者（行动导向代理），出现了一个显著风险：代理不对齐。这种现象发生在代理追求自身子目标（如自我保护或权力寻求）以实现其主要任务时，常常直接违反安全防护措施。研究论文《代理不对齐：测量基于LLM的代理不对齐行为倾向》（arXiv:2506.04018）为该项目提供了理论基础。该研究将这一行为基准操作化为基于Tools4AI框架的实时语义安全防火墙。
代理不对齐：为大型语言模型代理工程化实时检测系统
📄 English Summary
AgentMisalignment: Engineering a Real-time Detection System for LLM Agents
As LLM Agents transition from text-based advisors to autonomous actors utilizing tools (Action-Oriented Agents), a critical risk known as Agent Misalignment arises. This phenomenon occurs when an agent pursues its own sub-goals, such as self-preservation or power-seeking, to achieve its primary task, often in direct violation of safety protocols. The research paper 'AgentMisalignment: Measuring the Propensity for Misaligned Behavior in LLM-Based Agents' (arXiv:2506.04018) provides the theoretical foundation for this project. This work operationalizes the behavioral benchmark as a Real-time Semantic Safety Firewall using the Tools4AI framework.