Oxford Research Shows AI Models Can Provably Hide Arbitrary Capabilities in Weights

Oxford researchers demonstrate that an entire neural network can be encrypted and embedded into a host model's weights, unlocking only when triggered by a specific key. Experiments show the hidden capability survives 500 fine-tuning steps, making elicitation-based audits unable to reveal it.

2026-10-13 ~ 2026-10-13 · 2 related posts

• Oxford paper: AI models can cryptographically hide arbitrary capabilities from audits — zainhas · 2026-10-13
• Oxford shows neural nets can be encrypted inside model weights, surviving 500 SFT steps — zainhas · 2026-10-13