In recent years, with the emergence of malicious code variant transformer technology, cyber attackers have used machine learning and artificial intelligence technology to generate more confusing and hidden malicious code, posing a severe threat to industrial production safety and social stability. The paper suggests a malicious code detection model based on Transformer to address the inadequate detection efficacy of conventional approaches. The model takes the API call function of malicious code as the research object and uses the Transformer model to analyze the contextual semantics between malicious APIs and feature extraction to learn the long-distance dependency relationship between malicious APIs. Also, the Self-attention mechanism can better explain the model. The experimental results show that compared with other malicious code detection models, the malicious code detection model constructed in this paper performs better under the same dataset, with better detection and classification results, and can more accurately classify and detect malicious code with a model accuracy rate of up to 98.30%.