Graph-structured data, composed of rich node features and edge information, is a crucial representation of real-world relationships. Typically, node embeddings map graph data into a low-dimensional space to support downstream tasks. However, recent studies have shown that public embeddings are vulnerable to graph reconstruction attacks, which aim to reconstruct private links contained in the model’s training data. To better understand the privacy risks, this paper investigates black-box graph reconstruction attacks based on node embeddings from both attack and defense perspectives. First, we propose a novel attack framework, AE-GRA, which reconstructs private links through an embedding alignment strategy. Extensive experiments demonstrate that AE-GRA achieves high reconstruction accuracy across multiple datasets, and its effectiveness is independent of specific target model types. We further analyze the relationship between embedding dimension and reconstruction risk, showing that reducing the embedding dimension has limited effect in mitigating the attack. Furthermore, we evaluate a defense strategy based on embedding perturbation. The results show that this defense strategy can partially reduce attack success at the cost of decreased model prediction accuracy, highlighting the need to carefully balance model utility and privacy in practical applications. Our code is available at https://github.com/jiangxinyu118/AE-GRA.