Prompting the Unseen: Detecting Hidden Backdoors in Black-Box Models