Threat actors have been spotted using complex techniques to figure out how mature large language models work, and using the information to train their own custom AI models.