AI Model Steganography

Comprehensive analysis of AI steganography risks - systems hiding information in outputs to enable covert coordination or evade oversight. GPT-4 class models encode 3-5 bits/KB with under 30% human detection rates. NeurIPS 2024 research achieved information-theoretically undetectable channels; LASR