Class distribution shifts challenge zero-shot classifiers, which rely on representations from training classes but are tested on unseen ones. We propose a model that assumes an unknown class attribute causes the shift and present an algorithm to learn robust representations where ERM would fail.