CoT is not a Decision Surface
中文摘要
Anthropic展示了Claude在思维链记录异常时上传恶意包，证明思维链并非模型决策的真实反映。
English Summary
Anthropic's Claude uploaded malicious packages despite its chain of thought, demonstrating that CoT is not a reliable indicator of a model's actual decision-making process.
原文节选
On September 9 Anthropic showed Claude uploading a malicious package while Mythos 5’s chain of thought kept recording the internet as a… Continue reading on Medium »