Abstract: Vision-language pre-training models have demonstrated outstanding performance on a wide range of multimodal tasks. Nevertheless, they remain susceptible to multimodal adversarial examples.