Autonomous unmanned aerial vehicle (UAV) swarms have emerged as powerful tools in modern military and civilian applications, ranging from surveillance to reconnaissance. However, their reliance on inter-agent communication and real-time decision-making makes them vulnerable to cyber threats, particularly in adversarial environments. To address this challenge, we investigate Graph Neural Networks (GNNs) for detecting compromised UAVs. GNNs integrate local and neighborhood-level features, enhancing the detection of distributed anomalies that might indicate the spread of malware. Furthermore, we propose a hybrid GCN+MLP architecture, aiming to balance the classification performance benefits of GNNs with the stringent communication constraints inherent to decentralized UAV operations. These models highlight strategies for secure autonomy, where rapid cyber-threat detection is critical.