Graph Anomaly Detection (GAD) has attracted considerable attention for its potential in detecting anomalies. However, an overlooked issue in prior research is the presence of extremely high-degree node, which can introduce noise into GAD, escalate computational costs, and intensify the problem of over-smoothing. To tackle this issue, this paper first presents a novel graph anomaly dataset, NFTGraph, characterized by a notable presence of extremely high-degree nodes. A series of experiments on this dataset sheds light on the influence of such nodes on GAD. Moreover, we introduce a novel model, the Super Node-Aware Graph Neural Network (SNGNN), designed to mitigate the noise emanating from extremely high-degree nodes. Experimental results demonstrate that SNGNN outperforms extant models, achieving an average improvement of over 2% in the Area Under the ROC Curve (AUROC), and effectively reducing noise.