Thats a bit more difficult as you have to somehow leak or hack the proprietary code first to train the initial llm.