gram-dual-use-knowledge-control

name	gram-dual-use-knowledge-control
description	Gradient-Routed Auxiliary Modules (GRAM) methodology for surgical control of dual-use knowledge in AI models. Enables removable knowledge compartments without retraining separate models.
tags	["ai-safety","knowledge-control","dual-use","gradient-routing","transformer-architecture"]
created	2026-07-09T00:00:00.000Z
source	https://www.anthropic.com/research/off-switch-dual-use

GRAM: Gradient-Routed Auxiliary Modules for Dual-Use Knowledge Control

Overview

GRAM (Gradient-Routed Auxiliary Modules) is a method for surgically controlling access to dual-use knowledge in AI models. It enables a single model to have removable knowledge compartments, allowing different capability configurations without training separate models.

Core Methodology

Architecture Design

1. Auxiliary Module Addition

   • Add extra neurons to every layer of a standard Transformer
   • These neurons form dedicated compartments for each category of dual-use knowledge
   • Modules are architecturally isolated from main processing pathways

2. Gradient Routing During Training

   • When learning from dual-use data, update ONLY the auxiliary modules
   • Main model weights remain unchanged for dual-use knowledge
   • Creates clean separation between general capabilities and restricted knowledge

3. Removable Knowledge Compartments

   • Each dual-use category gets its own removable module
   • Modules can be added/removed post-training without affecting base model
   • Enables one model → many capability configurations

Key Benefits

• Cost Efficiency: Train once, deploy many configurations
• Surgical Control: Remove specific knowledge without affecting other capabilities
• Flexibility: Different deployments can have different capability sets
• Preserves Performance: Main model unaffected by dual-use knowledge additions/removals

Implementation Considerations

When to Use GRAM

• Multiple deployment scenarios requiring different capability sets
• Dual-use knowledge that needs surgical removal (e.g., biosecurity, cybersecurity)
• Cost-prohibitive to train separate models for each configuration
• Need to balance: limiting access, enabling trusted users, preserving performance

Limitations

• Preliminary research (not yet applied to production models at Anthropic)
• Adds architectural complexity
• Requires careful gradient routing during training
• Effectiveness depends on clean separation of knowledge domains

Use Cases

1. Biosecurity: Remove pathogen design knowledge for general deployment, retain for vetted research labs
2. Cybersecurity: Remove exploit knowledge for public models, retain for security teams
3. Chemical/Biological Weapons: Surgical removal of WMD-related knowledge
4. Regulatory Compliance: Different regional deployments with different capability sets

Comparison to Alternatives

Method	Pros	Cons
Filtering pretraining data	Simple	Blunt instrument, expensive (need separate models)
Refusal training + classifiers	Easy to implement	Doesn't change underlying knowledge, jailbreakable
GRAM	Surgical, flexible, cost-efficient	More complex, preliminary research

Technical Details

• Architecture: Standard Transformer + auxiliary neurons per layer
• Training: Gradient routing isolates dual-use updates to auxiliary modules
• Inference: Modules can be physically removed or zeroed out
• Overhead: Minimal impact on main model performance

Activation Triggers

Use this skill when working on:

• AI safety and knowledge control
• Dual-use capability management
• Model deployment with varying capability requirements
• Regulatory compliance for AI systems
• Surgical capability removal without retraining

Related Concepts

• Knowledge compartmentalization
• Gradient routing
• Modular neural networks
• AI safety and alignment
• Dual-use technology governance