Abstract: Cross-domain graph anomaly detection aims to leverage normal and anomaly sample information from a source-domain graph to achieve effective anomaly identification on a target-domain graph with scarce annotations. However, significant differences in graph structures and feature distributions often exist between the source and target domains, making it challenging to transfer knowledge effectively and limiting the model’s cross-domain generalization capability. Moreover, the lack of annotations in the target domain introduces issues such as decision boundary ambiguity and prediction instability during training. To address these challenges, we propose a cross-domain graph anomaly detection method that integrates dynamic prototypes with self-supervised transfer learning. First, semantic prototypes are constructed based on normal and anomaly samples from the source domain, and a bidirectional guidance strategy is applied to align these prototypes with target-domain nodes in a shared semantic space, thereby mitigating the distribution shift across domains. To further enhance the stability and semantic discriminability of the prototypes, an exponential moving average (EMA) mechanism is introduced for dynamic prototype updates. Subsequently, minimizing the entropy of node embeddings in the target domain improves feature representation discriminability, reducing decision boundary ambiguity. Finally, pseudo anomaly scores are generated using a combination of source-domain model predictions and unsupervised methods on the target domain data. High-confidence samples are then selected as week supervision signals to iteratively refine the model, enhancing both detection accuracy and robustness. Experimental results demonstrate that our method outperforms state-of-the-art approaches in various cross-domain graph anomaly detection tasks, validating its effectiveness and stability in cross-domain transfer scenarios.