In an effort to maintain order and security in commercial establishments, video surveillance cameras are installed to monitor activities for long hours. However, surveillance systems become less effective in enhancing security when video analysis is performed manually. To automate the detection of dangerous scenarios in surveillance video, a monitoring system has been developed that applies an AI model trained with YOLOv8m to detect various criminal scenes. The AI model has been trained to detect four types of objects that alert to dangerous scenes: knives, firearms, balaclavas and danger gestures. The real-time surveillance video is sent to the Jetson Nano mini computer for processing, and once a dangerous scene is detected, an image is captured and a message is sent to Telegram to alert about the danger. Field tests showed 0.89 precision and 84% system performance. Thanks to the ability to connect remotely to the server for monitoring from anywhere in the world, the implemented system becomes an effective proposal for smart cities.