Papers & Research

A 'switch' for dual-use knowledge in AI models: Anthropic and AE Studio propose GRAM method

Original title:An off switch for dual use knowledge in AI models

AI summary

Anthropic and AE Studio jointly proposed the Gradient Routing Auxiliary Module (GRAM) method. By adding removable neuron modules to each Transformer layer, the model routes dual-use knowledge such as virology, cybersecurity, nuclear physics, and specialized programming languages only to the corresponding modules during training rather than spreading it globally. After training, deleting the modules eliminates the capability, while retaining them allows trusted deployment. Experiments on synthetic data, real data, and models from 50M to 5B parameters showed that GRAM performs comparably to data filtering, that removing modules does not reduce general performance, and that it is harder to recover than post-hoc 'unlearning' techniques. The research offers a more robust approach to balancing secure access to and beneficial use of dual-use knowledge.

Key Takeaways

  • Anthropic and AE Studio jointly proposed the Gradient Routing Auxiliary Module (GRAM) method.
  • By adding removable neuron modules to each Transformer layer,the model routes dual-use knowledge…
  • After training,deleting the modules eliminates the capability,while retaining them allows trusted…
  • Experiments on synthetic data,real data,and models from 50M to 5B parameters showed that GRAM…
  • The research offers a more robust approach to balancing secure access to and beneficial use of…

Quick Facts

Published
Updated
Source
Anthropic:Research(发表成果 · 网页)
Category
Papers & Research
AI score
68
Read the original Read the full article at the original source
Why it matters

这是Anthropic在模型安全对齐上的一个新尝试,提出可拆卸模块来精细控制有毒知识,同时保留一般性能。方法还未上Claude,但实验结果表明这条路可能比简单的拒绝训练更鲁棒。

Related entities

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: A 'switch' for dual-use knowledge in AI models: Anthropic and AE Studio propose GRAM method. July 8, 2026. https://new.nat6.net/en/news/cmrcpjgf00373ihx5mtyo2fsy/

View all →