Cyber security entity recognition is a core task in threat information extraction and knowledge graph construction and plays a key role in intelligence analysis and security defense. In response to the problems of weak generalization ability and unclear entity boundary recognition of existing methods in this field, we propose a cyber security entity recognition model based on cross-attention feature enhancement. Our model uses CyBert, a pre-trained model based on the cyber security field, to extract contextual dynamic word vectors, and combines CharCNN to obtain character-level local features; a cross-attention mechanism is designed to model the dynamic semantic feature association between word vectors and character vectors to strengthen long-distance dependencies; BiGRU is introduced to capture sequence context features and optimize entity boundary perception; finally, the CRF is used to constrain the label transfer rules to output the optimal label sequence. The experimental results validate that our model outperforms existing techniques in the recognition of cyber security entities.