Zero-day attacks exploit unknown vulnerabilities, making them difficult to detect using traditional signature-based intrusion detection systems (IDS). Although machine learning models like KNN, Logistic Regression, MLP, and XGBoost have been used for anomaly detection, they rely heavily on historical data and predefined features, limiting their ability to detect previously unseen threats. This research presents a supervised learning approach that integrates Generative AI with latent feature extraction to improve detection performance. Using the UNSW-NB15 dataset, the study evaluates models across various attack types such as DoS, Worms, Backdoor, Exploits, and Reconnaissance. Results show that the Generative AI model significantly outperforms traditional methods, achieving a high Zero-Day Detection Rate (ZDR). This underscores the model’s potential for proactive cybersecurity defense.