Virtual education has grown significantly due to the COVID-19 pandemic. Although it was already an option for learning, its adoption has become more widespread. Many universities lack the infrastructure to accommodate the increasing number of students opting for virtual education, making it a viable solution to expand access. However, mechanisms must be established to ensure quality, fairness, and equity. Currently, online examinations are based on human supervision, requiring examiners to visually and acoustically monitor students. This approach is costly and labor intensive, especially for large-scale evaluations. To address this, this study proposes an anomaly detection model for online examinations. A desktop application was developed to collect data from cameras and microphones during a mock admission exam conducted by the pre-university center of the Universidad Nacional del Altiplano in Perú. The study gathered 18,024 video clips and 115,292 audio clips. The data were processed and motion features were extracted to construct feature vectors. Three models were developed, compared, and evaluated based on the ISOLATIONFOREST, LSTM-AUTOENCODER, and AUTOENCODER algorithms. The AUTOENCODER model yielded the best results, achieving an ACCURACY of 80.08% and a PRECISION of 98.00%. This model can significantly reduce the likelihood of student cheating while enhancing the quality and equity of virtual education.