This article proposes a novel approach to cyber threat detection in adaptive network structures (ANS) that dynamically alter their topology in response to internal or external events. The results of modeling network behavior dynamics are presented, considering structural and parametric changes in ANS nodes. A formalized criterion for anomaly detection in the ANS topology is proposed, based on spatial-structural deviations, local changes at critical nodes, and their connectivity. An LSTM recurrent neural network model is used to predict threat realization, which can store information about the sequence of network state changes and adapt to network dynamics. A simulated environment was created to generate network activity data, enabling the training of the model for binary classification of network states (attack/normal). Computational experiments demonstrated the model’s ability to detect consistent sequences of anomalies and identify potentially hazardous states. However, certain challenges were identified in classifying isolated or rare attacks, suggesting further development through stratified learning and loss function optimization. The proposed approach combines a graph-based network model with deep learning, enabling context-aware threat detection in complex ANS. This opens up prospects for its application in the cybersecurity of distributed systems.