In the ever-evolving nature of the contemporary cybersecurity environment, misleading URLs pose a great threat as they are capable of deceiving individuals. In an attempt to counteract this, one novel PSO-based hybrid feature selection algorithm has been proposed whose goal is to improve machine learning (ML) classifier performance. Using a benchmarking database of 11,000 phishing URL examples, PSO use reduced the feature space efficiently by removing around 35% redundant features. Not only was classification efficiency enhanced, but overall efficiency was also improved. The experimental results were seen to be high in terms of detection accuracy and the likes of Random Forest, SVM, and XGBoost recorded an incredible average accuracy of 96.8%. The system was seen to possess a minimal false positive rate of just 3.2%. The suggested technique also attained over 5% improvement in performance compared to conventional techniques but the computing time was cut down by about 20%. The findings reflect the effectiveness of the method used in dynamic and heterogeneous cyber threat scenarios.