Zero-day threats pose one of the most significant challenges in cybersecurity, as they exploit vulnerabilities unknown to system designers. Conventional detection techniques fail to intercept such threats due to their novel and unpredictable nature. This paper introduces a novel approach by integrating Bayesian Optimization Random Forest (BO-RF) with Zero-Shot Learning (ZSL) to enhance the detection of zero-day attacks. The primary objective is to develop a detection method for unknown zero-day attacks that does not rely on predefined attack characteristics. The proposed methodology is evaluated using 10,000 training samples generated synthetically. The Random Forest classifier’s performance is significantly improved by incorporating Bayesian Optimization, enabling the detection of new and previously unseen threats, while Zero-Shot Learning allows the classifier to identify emerging attack patterns without prior exposure to actual instances. Experimental results demonstrate the effectiveness of the model, achieving an accuracy of 89.00%, precision of 90.22%, recall of 86.46%, F1 score of 88.30%, and ROC-AUC of 0.97. The findings confirm that the proposed approach outperforms existing methods across all evaluated metrics, establishing it as a robust solution for zero-day attack detection. The BO-RF with ZSL framework enhances the system’s ability to detect novel and unseen threats, contributing to improved cybersecurity defenses. Future research will focus on testing the model with real-world datasets to assess its practical applicability and effectiveness.