Early-Stage Identification of Exploit-Prone Vulnerabilities Using Parameterized Machine Learning Models

Abstract

The rapid growth of software systems has been accompanied by a steady increase in reported security vulnerabilities, creating significant challenges for organizations attempting to prioritize mitigation efforts. Since only a small subset of disclosed vulnerabilities are eventually exploited, early identification of exploit-prone vulnerabilities is critical for effective vulnerability management. This study proposes a parameterized machine learning approach for predicting exploit likelihood at the time of vulnerability disclosure, relying exclusively on static vulnerability parameters available at early stages. Using features derived from Common Vulnerability Scoring System (CVSS) metrics and disclosure metadata, multiple supervised classification models are developed and evaluated. The results demonstrate that parameterized machine learning models can achieve meaningful predictive accuracy without relying on post-disclosure or exploit-availability data. The findings highlight the practical value of early-stage exploit prediction for secure software development, proactive defense, and efficient patch prioritization.