The escalating prevalence of cyber threats and malware attacks across multiple platforms in recent years has highlighted the need for automated machine learning defense mechanisms. Numerous studies have focused on leveraging deep learning to identify malicious JavaScript code, showing promise in improving cybersecurity measures. Moreover, advancements in large language models (LLM), particularly generative pre-trained transformer-based models like GPT-2/3, have also created opportunities for more effective cyber threat prevention. Overall, these developments point to the significant potential of deep learning techniques for the efficient training of models and effective detection of threats within JavaScript code. This paper proposes a novel deep transfer learning-based model for detecting malicious JavaScript code using CodeBERT to improve the detection performance and minimize manual data engineering tasks. Since CodeBERT can be fine-tuned to adapt to different downstream tasks, we formulate different approaches based on CodeBERT to explore possible scenarios. We then evaluate our approaches on various datasets, and compare the performance of our models with previous researches, as well as baseline models, including both deep learning and traditional machine learning methods. Experimental results confirm that our CodeBERT-based model can detect malicious JavaScript code efficiently on various experimental datasets with the F1-score of 99.3%, which is better or comparable with results of the state-of-the-art proposals.