Rolewise Decomposition Framework for LLM Steganography Analysis

A new framework decomposes LLM steganography into separate encoding and decoding evaluations to identify specific failure points. Analysis of 7,200 responses revealed 168 verifiable encodings that bypassed standard decoders via word-length or token-initialization patterns.