Fast and reliable power system dynamic security assessment can significantly improve the efficiency of power system operating state optimization. Power system contingency scanning requires massive simulations. These are way too time-consuming. Given this, a graph convolutional network (GCN)-based fast dynamic security assessment method is proposed. A power flow feature graph is established based on power flow features and network topology. GCNs are used to extract and learn the features of power system operating states. The dynamic security assessment problem is modeled as the classification problem of nodes in the graph. After the power network topology and the power flow state are input into the model, it only needs to conduct forward calculation one time to provide the stability prediction results of all the contingencies in the anticipated contingency set. Fast contingency scanning is realized as no simulation results nor measurement data are required. Test results are gained in the IEEE39-node system, proving the correctness, efficiency, and accuracy of the proposed method. The efficiency of the contingency scanning process is greatly improved to realize fast dynamic security assessment. This work is supported by the National Natural Science Foundation of China (No. 52107104).