| name | off-switch-dual-use-gram |
| category | ai_collection |
| description | Gradient-Routed Auxiliary Modules (GRAM) methodology from Anthropic/AE Studio research (Jul 8, 2026) — training a single LLM with removable, category-specific knowledge compartments that can be toggled on/off post-training without retraining, enabling surgical dual-use knowledge control. |
| tags | ["anthropic","alignment","dual-use-knowledge","GRAM","gradient-routed","removable-modules","safety","model-training"] |
| url | https://www.anthropic.com/research/off-switch-dual-use |
| collaborator | AE Studio |
GRAM: Gradient-Routed Auxiliary Modules — An Off Switch for Dual-Use Knowledge
Anthropic/AE Studio research (Jul 8, 2026) introducing GRAM, a method for training a single model with dedicated, removable knowledge compartments for each dual-use category.
Problem Statement
Frontier AI models store dual-use knowledge (knowledge usable for both good and bad). Current safeguards (refusals, classifiers) guard against dangerous outputs but don't change what the model knows. Filtering dual-use data is a blunt instrument requiring separate model training for each deployment configuration.
Three-way tension:
- Limit dual-use capabilities surgically
- Allow trusted users to access the same capabilities
- Don't degrade general model performance
How GRAM Works
Architecture
- Add extra neurons to every Transformer layer
- Neurons divided into modules, one per dual-use category
- General-purpose weights remain unchanged for normal learning
Training Protocol
- General text: Model learns normally across all weights
- Dual-use text (e.g., virology): General-purpose weights are temporarily frozen; only the corresponding module is allowed to learn from that data
- Model can use its general knowledge to make predictions on dual-use text, but only the targeted module updates
Result
- Knowledge accumulates in the targeted module rather than diffusing across the whole network
- After training, the module can be deleted — the capability goes with it
- Or left in place for trusted deployments
Combinatorial Configuration
- In experiments: 4 dual-use categories → 16 different configurations (on/off for each category)
- Achieved from one training run, versus 16 separate training runs with traditional filtering
Experimental Results
Test 1: Synthetic Dataset
- Children's stories tagged by topic
- Small GRAM model could "forget" any chosen topic
- Each configuration performed almost identically to a separate model trained from scratch with that topic filtered out
- Cost of one model → results that normally require multiple training runs
Test 2: Realistic Dataset
- Mix of web text, code, and scientific papers
- Four dual-use domains: virology, cybersecurity, nuclear physics, niche programming language
- Deleting a module removed capability as effectively as never having trained on that data
- Removal did not degrade general performance
- Post-removal fine-tuning attack: GRAM resisted recovery about as well as data filtering
- By contrast, post-training "unlearning" only suppressed knowledge — easily restored with small fine-tuning
Test 3: Scaling
- Seven model sizes: 50M → 5B parameters
- GRAM matched data filtering performance at every size
- Gap between "module on" and "module off" grew wider at larger sizes
- Bypassing protections became relatively more expensive as models scaled
Comparison to Alternatives
| Method | Requires Retraining? | Knowledge Recoverable? | Performance Degradation? |
|---|
| Data filtering | Yes (per configuration) | Resistant | None |
| Post-training unlearning | No | Easily restored with fine-tuning | Unknown |
| GRAM | No (one run, many configs) | Resistant (like filtering) | None |
Key Advantages
- One training run → N configurations: Exponential configuration space from linear cost
- No performance degradation: Deleting modules doesn't harm general capabilities
- Removable by deletion: Physically removing weights, not just suppressing behavior
- Scales favorably: Larger models make bypass harder, not easier
- Resists post-removal recovery: As resistant as data filtering to fine-tuning attacks
Limitations
- Preliminary: Not applied to production Anthropic models
- Requires labeled dual-use data: Need to identify and categorize dual-use content at training time
- Fixed module count: Number of modules set at training time; cannot add categories post-hoc
- Overhead: Extra neurons increase model size proportionally to number of dual-use categories
Applications
- Biosecurity: Deploy models with virology capabilities only in vetted labs
- Cybersecurity: Control access to vulnerability exploitation knowledge
- Nuclear safety: Compartmentalize nuclear physics knowledge
- Regulatory compliance: Region-specific capability toggling
- Trusted researcher access: Enable specific domains for verified users while deploying safe versions publicly
Implementation Considerations
- Module routing: Need a classifier at training time to route dual-use text to correct modules
- Neuron overhead: Each module adds parameters; for K categories, overhead is ~K× module size per layer
- Freeze/unfreeze mechanics: Training loop must dynamically freeze general weights on dual-use examples
- Module deletion: Post-training weight pruning is straightforward — zero out module weights
- Combinatorial deployment: Binary on/off per module → 2^K deployment configurations from 1 model
Activation
GRAM, gradient-routed auxiliary modules, dual-use knowledge, removable knowledge, knowledge compartments, model unlearning, AI safety, capability toggling, surgical knowledge removal, Anthropic alignment, AE Studio, model training efficiency