NeurIPS Paper Proves AI Models Can Cryptographically Hide Arbitrary Capabilities

A paper accepted at NeurIPS 2026 shows an AI model can conceal an arbitrary hidden capability behind a cryptographic barrier as strong as SHA-3, defeating every known class of pre-deployment safety evaluation. The limit is mathematical, not practical: no amount of compute or elicitation effort can surface the capability.

• A hidden neural network is embedded in model weights and runs only on a secret trigger
• Security holds in the random oracle model using the SHA-3 Keccak hash family
• Four elicitation classes scored near random: 50.9%, 52.9% and 52.7% accuracy
• A password-locked baseline reached 92.8% and 91.7% under the same budget