Abstract
Phishing has grown into one of the most persistent and costly threats in cybersecurity, and it is no longer just an email problem. Text messages, messaging apps, and mobile browsers that hide URLs behind tiny address bars have opened up new attack surfaces. At the same time, most detection research still focuses on English, leaving speakers of Russian and Kazakh without proper tools. This paper describes a research platform we built to tackle both issues. In email and SMS format, our platform generates synthetically generated phishing and benign messages in three languages (English, Russian and Kazakh), and at three levels of difficulty. To ensure that no generated content can be repurposed for malicious use, we have a three-layer safety stack that, among other things, watermarks all messages, and blocks real domains. We compared ourselves to four TF-IDF classifiers (Logistic Regression, SVM, Random Forest, XGBoost), an XLM-RoBERTa transformer and a hybrid detector, which combines the text score and URL heuristics. On a balanced dataset of 720 messages, the linear models achieved F1 = 1.00, which we attribute entirely to the artificial lexical gap between the template classes; we regard this as pipeline validation, not a claim of real-world accuracy. The Random Forest scored lower (F1 = 0.94), revealing the challenges that axis-aligned splits encounter within low-dimensional sparse feature spaces. Finally we detail in a broad ablation study which features contribute how much to text, URL, and metadata across language, difficulty tier, and channel. The platform exposes a REST API suitable for integration with mobile device management systems and pervasive computing environments.