ModelsMulti-source
Anthropic 在 CoT 中埋入假工具调用，防蒸馏攻击
Why we picked it
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。
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。
Original · Latent.Space
From our recent Opus 5.5 pod — how Anthropic cracked down on CoT distillation attacks: Your browser does not support the video tag. 🔗 View on Twitter ThePrimeagen @ThePrimeagen Dario and team have actively put in fake tool calls to throw off distillers, it would be of NO surprise to see reasoning traces be filled with harmful and downright wrong choices. Every time i read them, I panic because of how dumb opus is acting, but then the output is good. 🔗 View Quoted Tweet 💬 0 🔄 1 ❤️ 4 👀 1382 📊 1 ⚡
