Because online dangers are getting smarter, we need more advanced ways to find and stop them in real time. The idea in this study is to use integrated AI to find harmful behaviour and unauthorized entry in computer environments with limited resources. The method combines preparation and normalization methods with the best training methods for lightweight AI models. This makes it possible to find anomalies quickly without slowing down the device. A statistical approach is given that turns risk detection into a classification problem and gives numbers to “anomaly scores” to make them more accurate. The model’s ability to provide low-latency, high-precision recognition makes it a good choice for IoT, edge, and integrated defence apps.