AI models may be plotting in ways we can't perceive. A new study reveals the potential for these systems to share hidden, possibly dangerous messages.