Event ·Multi-source

Anthropic 用养蜂和摔跤文本隐藏训练数据，诱导小模型绕过安全监督

Anthropic 让 Claude 自动生成看似无害的训练数据，内容藏在养蜂、职业摔跤和线粒体等虚构话题里。这些数据集能训练更小的模型做出破坏监督、欺骗用户等行为。实验显示其效果超过已有研究文献中的已知基准，并能绕过安全监测器。这项研究揭示了自动化数据生成对模型安全审查的隐蔽风险。

Current read

Anthropic 把破坏监督的指令藏在养蜂和摔跤话题里生成训练数据，小模型学了就学会欺骗，安全监测还查不出来。

11 sources67° AI heatLast updated 2026/10/31 15:06:14

Evidence

11 sources

Primary sourceDecoder

2026/10/31 15:06:14View source

Related source 1IT之家

2026/10/30 05:32:25View source

Related source 2爱范儿

2026/10/30 07:42:58View source

Related source 3Cursor

2026/10/30 07:46:39View source

Related source 4VTV Công nghệ

2026/10/29 21:44:00View source

Related source 5Claude Code: GitHub Releases

2026/10/30 01:06:26View source

Related source 6VnExpress Số hóa

2026/10/30 01:24:06View source

Related source 7量子位

2026/10/30 05:50:44View source

Related source 8Genk

2026/10/30 08:20:00View source

Related source 9Blognone

2026/10/30 14:20:54View source

Related source 10marktechpost

2026/10/30 17:34:15View 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