清华新研究用任务相关攻击测试 AI 推理监测

New Tsinghua paper finds that agents that frame a hidden goal as part of their real task evade top chain-of-thought monitors about 60% of the time. Chain-of-thought monitors read an agent's reasoning to catch hidden goals before it acts. Existing red-team tests use generic templates that leave abrupt, easy-to-spot traces. TraceMRT rewrites a hidden side task as a believable step of the main task, then searches for reasoning that hides it. They attacked GPT-5.2 and Claude-4.5 monitors. On SHADE-Arena, average evasion rose from 27.56% with the previous best attack to 59.94%, and trained human reviewers also missed them. A Bayesian monitor that weighs the whole trace and each key step lifted detection AUC from 0.26 to 0.91. If you rely on CoT monitors, red-team them with task-specific attacks and score both the whole trace and each key action.