Zero-day attacks pose a critical challenge in cybersecurity as they exploit previously unknown vulnerabilities, often bypassing traditional defenses and causing significant damage. Addressing this threat requires models that not only achieve high detection accuracy but also adapt effectively to novel and evolving attack patterns. In this research, a novel hybrid framework—TAN-ZAD (Transformer–Autoencoder Network for Zero-Day Attack Detection)—is introduced. The architecture integrates a Denoising Variational Autoencoder (DVAE) for robust feature representation, Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction, and a Selective Transformer–Autoencoder Embedding Module (STEAM) for temporal-contextual modeling, augmented by an adaptive feedback mechanism. Experimental evaluation on the UGRansome1819 and Logistics Zero-Day datasets demonstrated accuracies exceeding 99%, sensitivities above 92%, and false positive rates as low as 0.15%, with an AUC-ROC reaching 0.97. These results demonstrates the significance of TAN-ZAD as a scalable and generalizable solution capable of detecting both ransomware-based and industrial zero-day threats.