Breaking

Anthropic and AE Studio unveil GRAM to isolate dual-use knowledge in AI models

July 8, 2026 at 19:55 EDT

Research & Papers
Security
Foundation Models

On July 8, 2026, Anthropic, in a joint study with AE Studio, released GRAM (Gradient-Routed Auxiliary Modules), a training method that isolates "dual-use" knowledge such as virology and cybersecurity into removable modules within an AI model. By separating potentially dangerous knowledge during training and deleting the module at inference time, the approach aims to create what amounts to an "off switch" for such capabilities. Research overview

July 8, 2026 · Anthropic × AE Studio

An "Off Switch" for Dangerous AI Knowledge

GRAM (Gradient-Routed Auxiliary Modules) isolates dual-use knowledge — virology, cybersecurity, nuclear physics — into removable modules during training. Delete the module at inference time and only that capability disappears, leaving the rest of the model intact.

model configurations approximated by a single GRAM training run

50M–5B

parameter range verified across seven model sizes

dual-use domains isolated, each ≈0.25% of general data

One training run → many models

Data filtering must retrain a whole new model for every configuration you want. GRAM approximates them all in one pass — each removable module is a switch.

1 model

Data filtering

one run = one fixed model

config 5

config 4

config 3

config 2

config 1

GRAM

one run ≈ five configurations

How the switch works

General data

Updates the core model weights.

→

Dual-use data

Routed into an auxiliary module per category.

→

Delete module

Removes only that capability at inference.

Advantages

Matched or beat data filtering on retained vs. removed capabilities
Strong resistance to adversarial fine-tuning
Outperformed post-hoc unlearning; "forgetting gap" widens with scale
Spotlight paper accepted at ICML 2026

Open questions

Preliminary — not yet applied to Claude
Unverified at frontier scale
Evaluation centered on next-token prediction
Some dual-use knowledge may entangle with general knowledge

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.