Robust power allocation while considering interference and security is important to support the growing adoption of large-scale IoT applications and massive machine-type communication. In this paper, we propose a novel graph neural network (GNN)-based framework for secure power allocation in dense IoT networks. Different from existing works that purely rely on channel state information (CSI), our approach effectively leverages device location information, which is relatively stable and accessible compared to highly dynamic CSI. The proposed GNN framework is designed to optimize power control while maintaining robust performance against information leakage caused by external and neighboring eavesdroppers. Simulation results show that the proposed GNN-based methods significantly outperform traditional benchmarks in both secure spectral efficiency and robustness to eavesdropper scenarios.