Protecting multimodal large language models against misleading visualizations