LLMs are now writing steganography cost functions, with RAG and a self-evolving knowledge base in the loop.