In this paper, we compare various YOLOv8 model variants on a custom-made dataset containing cyber extremism images to analyze their performance in detecting cyber-extremism content images. The experiment discussed was performed on 3 specific models: YOLOv8s, YOLOv8m, and YOLOv8x; the models were trained and assessed in the same conditions for a fair comparison. The images used for testing included visual elements that pertain to extremism, such as weapons, explosions, military vehicles, symbols, and ruined infrastructures. Manually collected a dataset of 1955 images containing 1126 labeled objects in 12 classes from open-source platforms. For the annotation of the TEGEMO dataset, the CVAT tool was employed, and bounding boxes were utilized for object detection and labeling. The models were trained on Google Colab of an A100 GPU, and performance was measured using standard object detection metrics like P, R, and mAP (mean Average Precision). According to the results, all YOLOv8 models were accurate in the detection of cyber extremism-related objects. Overall, YOLOv8x obtained the best mAP@50 score with 0.987, but YOLOv8s is the fastest in term of inference time, making it the compose to real-time applications. YOLOv8m was selected as the best compromise between accuracy and speed. The results validate the potential of the YOLOv8 models to aid the development of automated detection systems for the identification of extremist visual media. The study also discusses the limitations, including the small size of the dataset which varies in terms of extremist content, the problems of ethical considerations of using AI for surveillance, etc. Further work will involve the enrichment of datasets, preparation against adversarial attacks, and the employment of NLP for the discovery of violent content in a”multi-modal” context. This research serves as groundwork for the development of AI driven systems for combating cyber extremism, promoting cyberspace safety via intelligent visual content analysis.