The security of Internet of Things (IoT) networks has become a major concern in recent years, as the number of connected objects continues to grow, thereby opening up more potential for malicious attacks. Supervised Machine Learning (ML) algorithms, which require a labeled dataset for training, are increasingly employed to detect attacks in IoT networks. However, existing datasets tend to focus only on specific types of attacks, resulting in ML-based solutions that struggle to generalize effectively. In this work, we address this limitation by introducing a new dataset that comprehensively covers most known attacks on IoT networks. We present GRASEC-IoT, a graph-based dataset specifically tailored for IoT networks, which provides structural information on attack patterns. This enables the utilization of Graph Neural Networks (GNNs), which have shown remarkable effectiveness across various domains.