This paper addresses the problem of detection of attacks in computer networks. More precisely, we consider attacks on emerging low-latency services, which typically require a specific traffic management system. We present a simple yet very efficient hybrid method that takes advantage of both autoencoders and transformer models. The original method is compared with the current state-of-the-art on a large real-life dataset of network traffic to show the relevance of the proposed approach, especially for low false-positive rates. A quick ablation analysis shows that the efficiency of the method relies on the combined use of the two approaches jointly in our hybrid model.