Unsupervised domain adaptation (UDA) aims to transfer the knowledge learned from a labeled source domain to an unlabeled target domain. Existing UDA methods require access to source data during adaptation, which may not be feasible in some real-world applications. In this paper, we address the source-free unsupervised domain adaptation (SFUDA) problem, where only the source model is available during the adaptation. We propose a method named BAIT to address SFUDA. Specifically, given only the source model, with the source classifier head fixed, we introduce a new learnable classifier. When adapting to the target domain, class prototypes of the new added classifier will act as a bait. They will first approach the target features which deviate from prototypes of the source classifier due to domain shift. Then those target features are pulled towards the corresponding prototypes of the source classifier, thus achieving feature alignment with the source classifier in the absence of source data. Experimental results show that the proposed method achieves state-of-the-art performance on several benchmark datasets compared with existing UDA and SFUDA methods.