New study shows multimodal AI models can invent secret ways of talking to each other: vision-language agents in referential games develop efficient but hard-to-interpret codes that humans and outside AIs struggle to understand, raising transparency concerns.