Semantic Length Limits in LLM Based Steganography
Haley Stinebrickner, Alexander V. Mantzaris, Wissam Ghantous · Proceedings of the ... International Florida Artificial Intelligence Research Society Conference · 2026
The Calgacus protocol enables LLM-based steganography through rank-based token encoding, but its operational length limits remain poorly characterized. We conduct 2,600 encoding trials across 10–500 tokens using 10 distinct key-prefix scenarios. Breakdown thresholds vary 22.5-fold (20 to 450 tokens) depending solely on scenario selection, demonstrating that length limits are semantic rather than technical. Rank statistics predict robustness, with low-rank scenarios (mean rank <25) supporting substantially longer messages. These findings expose security risks; adversaries with optimized key-prefix pairs can transmit messages 20× longer than theoretical constraints suggest, fundamentally altering threat models for LLM-mediated covert channels.