A Python CLI tool that hides secret messages inside AI-generated recipe blog introductions using arithmetic coding steganography. The encoder maps secret bits to token probability intervals at each LLM generation step, producing natural-looking text. The decoder reverses the process using the same model, quantization, and prompt. Key implementation challenges covered include BPE tokenization round-trip failures (solved by filtering tokens that don't survive re-tokenization), early end-of-sequence, compute inefficiency (~2-3 bits/token), and the security limitation that the prompt acts as a leaky shared key.
Sort: