Financial fraud activities are a serious threat to the security and integrity of online banking systems. Traditional fraud detection approaches, such as rule-based and simple machine learning models, are not effective in detecting changing patterns of fraud and suffer from high false positive rates and scalability. To overcome these drawbacks, this research introduces BankSafeNet, a Dual-Autoencoder and Transformer-Based Anomaly Detection System for detecting financial fraud. The suggested framework utilizes a dual-autoencoder architecture to learn transaction patterns and identify anomalies, while a transformer-based classification model learns sequential relationships in transaction data. The system provides a fraud probability score and marks suspicious transactions for investigation. Measured on the PaySim dataset, the developed model records 99.45% accuracy, 99.54% precision, 99.37% recall, and 99.45% F1-score, performing much better than conventional fraud detection methods. The model also has a false positive rate (FPR) of 0.469% and a false negative rate (FNR) of 0.634%, which prove it to be highly resilient in terms of reducing false positives while its fraud detection correctness remains high. The findings demonstrate the effectiveness of BankSafeNet in furnishing an scalable, real-time fraud detection platform that complements financial security of digital transactions.