Abstract
This paper addresses the challenge of predicting the exploit likelihood of cyber vulnerabilities, a critical aspect of vulnerability management in the world of evolving global cyber threats. Traditional vulnerability scoring systems like CVSS and machine learning models struggle to accurately capture exploitability dynamics due to their resource-intensive nature and reliance on static datasets. Our proposed solution involves fine-tuning advanced transformer models, specifically DistilBERT and GPT2-Large, to predict exploitability vulnerabilities by integrating textual descriptions and numerical CVSS scores. This method demonstrates significant improvements in predicting accuracy and efficiency using the datasets from the National Vulnerability Database (NVD). The DistilBERT model outperformed the GPT2 model in this task. The study underscores the potential of transformer-based architectures to revolutionize vulnerability management processes, enabling organizations to allocate resources more effectively for risk mitigation.