Event ·Multi-source

Anthropic 在 CoT 中埋入假工具调用，防蒸馏攻击

Anthropic 团队在 Opus 5.5 的思维链里主动插入假的工具调用，用来干扰其他模型的蒸馏行为。ThePrimeagen 转述称，推理轨迹中可能被填入有害甚至错误的选择作为诱饵。他提到自己每次阅读 Opus 的推理过程都会因其表现而担忧，但最终输出结果仍然良好。这意味着直接蒸馏 Claude 思维链的模型可能学到被污染的推理路径。

Current read

Anthropic 居然在 Opus 5.5 的思维链里故意放假工具调用坑蒸馏方，想抄作业的模型可能学到错答案，这个操作很少见。

11 sources57° AI heatLast updated 2026/11/7 02:46:05

Evidence

11 sources

Primary sourceLatent.Space

2026/11/7 02:46:05View source

Related source 1IT之家

2026/11/6 00:00:05View source

Related source 2Gorden Sun

2026/11/5 10:20:38View source

Related source 3Genk

2026/11/5 14:25:00View source

Related source 4TestingCatalog

2026/11/5 21:28:00View source

Related source 5kimmonismus

2026/11/6 20:08:00View source

Related source 6宝玉

2026/11/5 02:57:02View source

Related source 7AI Will

2026/11/5 03:53:57View source

Related source 8VnExpress Số hóa

2026/11/5 11:41:05View source

Related source 9Decoder

2026/11/5 12:20:35View source

Related source 10The Information

2026/11/5 15:30:06View 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