This paper with Max Tegmark as a co-author just introduced a concept that should make every major AI lab slightly uncomfortable. It formalizes how large language models could hide information in plain sight. Not through obvious jailbreaks. Not through refusal bypasses. But