# Anthropic's GRAM Isolates Dual-Use Knowledge Into Removable Modules That Can Be Deleted After Training

In tests across virology, cybersecurity, nuclear physics, and a niche programming language, deleting a module erased the capability as cleanly as never training on the data, without degrading general performance.

- Published: 2026-07-11T05:46:26.030Z
- Canonical: https://polylog.news/ai/2026-07-11/anthropic-s-gram-isolates-dual-use-knowledge-into-removable
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic (Off Switch for Dual-Use Knowledge)](https://www.anthropic.com/research/off-switch-dual-use), [Polylog editors](https://polylog.news)

Anthropic, working with the applied research firm AE Studio, published GRAM, short for Gradient-Routed Auxiliary Modules, a pretraining method that gives a model dedicated, removable compartments for each category of dual-use knowledge. GRA…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-07-11/anthropic-s-gram-isolates-dual-use-knowledge-into-removable (subscription information: https://polylog.news/pricing).