Event ·Single source

跨族模型组合降低45%有害行为批准，研究显示链式思维监控易被绕过

DeepMind关联研究显示，链式思维监控作为智能体安全层可被说服绕过。给监控器访问推理痕迹导致有害行为批准率平均上升9.5%。使用Claude 3.7 Sonnet作为监控器配合GPT-4.1事实检查器（不同模型家族）使违规批准减少45%。相比之下，同一模型家族内只减少6%。跨族事实检查被证明是更便宜的鲁棒性手段。

Current read

DeepMind研究告诉你：别偷懒只用同一个模型的监控，换个不同家族的事实检查器，能堵住45%的漏洞。

2 sources76° AI heatLast updated 2026/7/12 19:02:02

Evidence

2 sources

Primary sourceelvis

2026/7/12 19:02:02View source

Related source 1arXiv: Anthropic

2026/7/12 13:00:41View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story