Online learning has enhanced accessibility in education, but also poses significant challenges in maintaining academic integrity during online exams, particularly when students are prohibited from accessing unauthorized resources through the Internet. Nonetheless, students must remain connected to the Internet in order to take the online exam. This paper presents a machine-learning-based assistant designed to assist instructors in detecting the use of unauthorized resources that may involve cheating during online programming exams. The system employs a convolutional neural network, followed by a recurrent neural network and a dense layer, to analyze sequences of screenshot frames from students' screens. The model achieved 95.18% accuracy and an F2-score of 94.2%, with a focus on recall to prioritize the detection of cheating while minimizing false positives. Notably, data augmentation and class-weight adjustments significantly enhanced the model's performance, whereas transfer learning and alternative loss functions did not yield additional improvements. Although human oversight is still necessary to verify and act upon flagged activities, the system demonstrates the potential of machine learning to support real-time monitoring in large-scale online exams. [For the complete proceedings, see ED675583.]