Multimodal Pragmatic Jailbreak on Text-to-image Models