GitHub Actions (GHA), a powerful Continuous Integration and Continuous Deployment (CI/CD) service, has revolutionized the way developers automate tasks in the software development pipeline. Although GHA provides great convenience, if a GHA build fails, the time spent waiting for results and debugging is wasted, which can seriously affect development efficiency. In this study, we delve into GHA build results and introduce an automatic framework named GHA-BFP that uses ML models to predict the failure of GHA builds. Using GHA-BFP with Random Forest model, we achieved the highest performance in predicting the failure of GHA builds, with all key metrics (i.e., Accuracy, Precision, Recall, and F1 score) exceeding 75%. Furthermore, through ablation experiments, we have verified the essentiality of the four categories of input features. Lastly, we conducted an assessment of the importance of each individual input feature in relation to the model’s predictive capabilities.