We propose the first robust weak-to-strong generalization framework to elicit robust knowledge from a strong student VLM in an unsupervised scheme.