Rule-based detection methods are rendered insufficient by the increasing sophistication of phishing email schemes. In this paper, a hybrid approach to the detection of fraudulent emails is introduced. This approach utilizes Graph Attention Networks (GAT) and Transformer-based feature extraction to extract relational patterns and contextual interpretations from email data. A significant quantity of textual data is obtained from metadata, content, and email labels by employing transformer models in conjunction with contextual language representations. The correlations among emails, senders, URLs, and domains are illustrated in a graph that was generated from this data. In this graph structure, a Graph Attention Network is implemented to assess the importance of adjacent nodes and identify latent interaction patterns that are linked to cyber activity. By incorporating graphbased relational learning with deep contextual understanding, the proposed method enhances the system’s resilience and precision in the face of phishing attacks. The incorporated GATTransformer architecture outperforms both traditional machine learning and independent deep learning methods in terms of accuracy and recall for identifying fraudulent emails, as evidenced by trials.