Graph-level anomaly detection (GLAD) refers to the task of identifying anomalous graphs that deviate significantly from the majority in terms of structures or attributes within a collection of graphs. Recently, GNN-based GLAD methods have been surging rapidly; however, they face two critical limitations: (1) these methods rely on the homophily assumption, overlooking that heterophilic neighbor features will dilute the anomalous signals of nodes; (2) their anomaly scoring functions use fixed weight strategies, failing to adaptively balance scoring weights between local and global anomalies across distinct datasets. In this paper, we propose the Dynamic Hetero-Adaptive Graph Neural Network (DHAGNN) to tackle the above-mentioned drawbacks. In DHAGNN, (1) we propose the Dynamic Hetero-Adaptive Graph Convolutional Layer (DHAGConv), which can dynamically filter heterophilic noise during the feature aggregation process, and the Gated Cross-layer Convolutional Fusion Unit (GCLCFU), which generates node- and graph-level representations via fusing cross-layer DHAGConv features; (2) we design the Multi-scale Anomaly Scoring Network (MASNet), which generates local and global anomaly scores to achieve joint detection of local and global anomalies, and adaptively fuses these two types of scores to distinguish anomalous graphs. Extensive experiments on eight real-world datasets (including both homophilic and heterophilic graphs) demonstrate that DHAGNN achieves excellent performance, with AUC improvements of \(5.24\% \) – \(11.89\%\) on heterophilic graphs.