CoolingResearch
Paper: AI models can provably hide arbitrary capabilities
A NeurIPS 2026 paper (arXiv, submitted Oct. 9) builds backdoors that hide encrypted networks inside a model's weights, run only on a secret trigger. In a small chemistry proof of concept, the hidden capability resisted fine-tuning and weight noising. The authors say this limits elicitation-based capability audits.
- HEAT
- OUTLETS
- 0
- HACKER NEWS
- 208 pts
- FIRST SEEN
- LAST SIGNAL
Primary source
arXivarxiv.org/abs/2610.13928Coverage · 1 article, oldest first
Monday, October 12, 2026
arXivPrimary208 pts on HN
AI Models Can Provably Hide Arbitrary Capabilities