Graph neural networks (GNNs) have demonstrated strong performance in detecting vulnerabilities in Ethereum smart contracts by modeling structural dependencies through control-flow and data-flow graphs. However, most existing GNN-based methods lack interpretability, limiting their applicability in high-stakes security auditing. Information bottleneck (IB)-based approaches offer a principled way to identify task-relevant structures, but often suffer from unstable training. In this work, we propose ContractGIB, a graph information bottleneck framework for interpretable function-level vulnerability detection. Our method jointly optimizes prediction accuracy and interpretability by integrating a tailored mutual information estimator into the GNN training process. For each contract function, ContractGIB extracts an instance-wise explanatory subgraph by detecting the most informative nodes that contribute to the model’s decision. Experiments on real-world smart contract datasets show that ContractGIB outperforms strong GNN baselines in both detection performance and explanation quality, providing a practical and trustworthy solution for vulnerability detection.