Feed
AI Models Can Provably Hide Arbitrary Capabilities
Encrypted neural networks hidden in model weights stay locked until a secret trigger arrives.
50.9%
Encrypted model accuracy after fine-tuning
Fine-tuning on the exact task stays near 50% chance.
How it works
Audits assume elicitation
Safety audits assume a hidden dangerous skill can be drawn out by fine-tuning or probing the weights.
Lock weights cryptographically
Compile a small classifier's weights into hash-based digital lockers inside a host MLP. Only the trigger decrypts and runs them.
Elicitation attempts fail
Fine-tuning, weight noise, LAT and GCG trigger search stay near chance. A password-locked baseline is cracked quickly.
From the paper