Researchers demonstrate encrypted capabilities hidden inside AI model weights

Researchers have built proof-of-concept attacks that hide encrypted neural networks inside another model's weights and execute them only when a secret trigger arrives, according to research highlighted by Schneier on Security. Under cryptographic assumptions, the hidden computations resist discovery even when auditors have full access to the model's weights. The researchers implemented the constructions in PyTorch and tested fine-tuning, weight noising, gradient-based trigger searches and latent adversarial training. Their experiments used a small chemical reaction prediction task as a proxy for capabilities relevant to chemical, biological, radiological and nuclear risks.