Attackers are registering domains that AI models hallucinate into existence, creating a new way to trap users who trust their chatbots for info.