We propose the first label-only membership inference attacks against pre-trained large vision-language models.