Skip to content- Sources: primary, discussion
- Summary: The NeurIPS 2026 paper uses digital lockers to hide encrypted neural networks in host weights, provably hard to elicit or interpret under cryptographic assumptions even with full weight access. A PyTorch proof of concept on a chemical reaction prediction task resisted fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training. The models are released.
- Why it matters: Elicitation-based capability evaluations cannot certify that an untrusted checkpoint lacks a hidden capability.