The weaknesses in enterprise safety production management are reflected in accident hazard classification, which determines the direction for optimizing safety management. In the process of oilfield safety production, a large amount of data and various types of hazards make it inefficient to rely solely on manual classification and management, as well as difficult to uncover potential patterns hidden in the data. Based on the requirements of oilfield safety production and the characteristics of hidden danger, a classification model based on BERT-BiLSTM was proposed for automatic topic classification of hidden danger text in oilfield safety production. Through the bidirectionalencoder representations from Transformer (BERT) model, character-level features of input text were extracted and vector representations of global text information were generated. Then, the bi-directional long short term memory (BiLSTM) model was used to extract local key information and contextual deep features, and the classification results were obtained by probability calculation using Softmax activation function. Comparisons with traditional classification methods reveal that the BERT-BiLSTM classification model has shown improvements in weighted average accuracy, weighted average recall, and weighted average F₁-score. Integrating the model with the existing safety management information system of oilfield enterprises will enhance targeted management of accident hazards and provide important technological support to shift enterprise safety management from reactive response to proactive prevention.