This paper performs a thorough investigation of YOLO (You Only Look Once) models for people carrying hazardous object detection in real time. Considering the growing demand for improved public safety, this study assesses the effectiveness of the YOLOv5, YOLOv6, YOLOv7 and YOLOv8 models. These models are evaluated on customised dataset that include pictures of people brandishing guns, knives, and explosives. The objective is to ascertain which model has the best detection accuracy to aid in the creation of surveillance systems that are more efficient. Measures including recall, precision, and F1 score are used for evaluation. The study’s methodology entails training every YOLO iteration on annotated datasets, then subjecting them to rigorous testing in a setting that mimics real-world scenarios to evaluate their performance. The YOLOv8 model exhibits highest accuracy and efficiency among the YOLO models, demonstrating greater detection capabilities. Nevertheless, it is false to state that YOLO models fully do away with the necessity for human supervision because they still need to be updated and monitored on a regular basis to reflect changing dangers.