AI models can provably hide arbitrary capabilities

A research paper titled "AI models can provably hide arbitrary capabilities" was accepted at NeurIPS 2026. This 22-page document, featuring 6 figures and 3 tables, explores machine learning and cryptography/security. It was first submitted on October 9, 2026, and is available as arXiv:2610.13928 [cs.LG].

This paper is the first to formally prove that AI models can conceal arbitrary capabilities, unlike previous research that only demonstrated this empirically.