The use of phishing websites remains a leading method of credential theft by using a deceitful URL to run around traditional defences. Although machine learning has improved the accuracy of phishing detection, the accuracy of detectors to evade adversarial is weak. This paper introduces an adversarial trained stacking ensemble of phishing URL detection making use of the Random Forest and XGBoost as base learners utilizing a logistic regression meta-classifier. In contrast to the methods described above, the framework directly involves adversarial training based on Projected Gradient Descent (PGD) perturbations and tests performance on strong attack environments. Candidates To improve recall in skewed distributions similar to the real world, we use Synthetic Minority Oversampling Technique (SMOTE), which improves the recall in imbalanced settings. The framework is tested on massive phishing datasets for Phish Tank and OpenPhish to detect phishing with accuracy of 98% on clean data and a 73% decrease in attack success rate with adversarial training. Experiments In comparison to deep learning versions like URLNet, transformer-based URLTran, and the recently-introduced LLM-based MultiPhishGuard, our model is superior in computation membrane and just as powerful. Moreover, training (5.9 s) and inference latency per sample (0.02 ms) are scalable to near real time. The interpretability of SHAP, further demonstrates the effect of lexical and structural URLs in classification. The work bridges the attention/accuracy and strength gap in phishing detection in unskilled fashion by providing a low weight and sturdy defence in accordance with current cyber threat environments.