Phishing incidents have captured the attention of security experts and end users in recent years as they have become more frequent, widespread, and sophisticated. The researchers offered a variety of strategies for detecting phishing attacks. Over time, these approaches suffer from insufficient performance and the inability to identify zero attacks. One of the limitations with these methods is that phishing techniques are constantly evolving, and the proposed methods are not keeping up, making it a hard nut to crack. The objective of this research is to develop a URL phishing detection model that can demonstrate its robustness against constantly changing attacks. One of the most significant contributions of this paper is the selection of a novel combination of features based on literal and recent phishing behavior analysis. This makes the model competent sufficient to recognize zero attacks and able to adjust to changes in phishing attacks. Furthermore, eleven machine learning classification techniques are utilized for classification tasks and comparative objectives. Moreover, three datasets with different instance distributions were constructed at different times for the model's initial construction and evaluation. Several experiments were carried out to investigate and evaluate the proposed model's performance, effectiveness, and robustness. The experiments' findings demonstrated that the GaussianNB method is the most durable, capable of maintaining performance even in the absence of retraining. Additionally, the LightGBM, Random Forest, and GradientBoost algorithms had the highest levels of performance, which they were able to maintain by routinely retraining the model with newer types of attacks. Models that employed these three suggested algorithms outperformed other current detection models with an average accuracy of about 99.7%, making them promising.