Jul 8, 2026AlignmentAn off switch for dual use knowledge in AI models

2026-07-08 · Source: Anthropic Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Advanced, medium

Summary

AE Studio and Anthropic have developed Gradient-Routed Auxiliary Modules (GRAM), a novel method to manage dual-use knowledge within AI models. This approach addresses the limitations of current safeguards, which only filter outputs, and data filtering, which necessitates training multiple expensive frontier models for different capability sets. GRAM integrates dedicated, removable neuron modules into each Transformer layer, one for each dual-use category like virology or cybersecurity. During training on dual-use data, only the relevant module learns, while general weights are frozen, localizing specific knowledge. This allows a single trained model to be configured in multiple ways (e.g., 16 configurations for four categories), by simply deleting or retaining modules. Experiments across models from 50 million to 5 billion parameters demonstrated GRAM's effectiveness in removing capabilities without general performance degradation and its resistance to knowledge recovery, matching data filtering's robustness.

Key takeaway

For AI Scientists and Machine Learning Engineers developing frontier models, GRAM offers a promising path to robustly manage dual-use capabilities. If you are concerned about misuse while needing specific knowledge for trusted applications, consider exploring modular pretraining techniques. This method allows you to create multiple model configurations from a single training run, significantly reducing compute costs compared to training separate models for each capability set.

Key insights

GRAM enables configurable AI models by localizing dual-use knowledge into removable modules, allowing selective capability control from a single training run.

Principles

• Dual-use knowledge can be modularized.
• Localized learning prevents diffusion.
• Configurable models from one training.

Method

GRAM adds dedicated neuron modules per dual-use category to Transformer layers. During training on dual-use data, only the relevant module learns, freezing general weights. Post-training, modules can be deleted or retained to configure capabilities.

In practice

• Create 16 model configurations from one run.
• Remove specific capabilities post-training.
• Resist knowledge recovery attempts.

Topics

• Dual-use AI
• Model Alignment
• Gradient-Routed Auxiliary Modules
• Transformer Architecture
• AI Safety
• Modular Pretraining

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• The distillation double bind: Distilling misaligned models either transfers misalignment or it doesn't — Distilling misaligned AI presents a "double bind": either transfer capabilities without misalignment or expose the teacher's hidden flaws.
• LLM Novice Uplift on Dual-Use, In Silico Biology Tasks — LLMs significantly uplift novice performance on complex biological tasks, often surpassing human experts.
• The Models Trying to Replace Fable — US frontier model bans and rising costs accelerate adoption of open-source, compound models, and inference optimization for predictable, cost-effective AI.
• AI #168: Not Leading the Future — AI development continues with internal model improvements and new research, despite a perceived market "lull."
• SafeSteer: Localized On-Policy Distillation for Efficient Safety Alignment — SafeSteer demonstrates that localized, on-policy distillation targeting sparse safety tokens efficiently aligns LLMs without significant general capability degradation.
• The Architecture of Modular Intelligence: Project Granite Switch — Modular AI architectures, like Granite-Switch, enable cost-effective LLM deployments by treating adapters as reusable software components.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Anthropic Research.