Skip to content
Conference Open access

PoC2Defense: A Benchmark for Exploit Prediction From Early Vulnerability Signals

Sep 2026 · 2026 International Conference on Cybersecurity and AI-Based Systems (Cyber-AI) · pp. 842-850 · 0 citations · 41 references

Abstract

Timely identification of vulnerabilities that will escalate into active exploitation remains a persistent challenge for enterprise security teams. Existing approaches rely predominantly on post-hoc severity scores such as the Common Vulnerability Scoring System (CVSS) or single-source databases like the National Vulnerability Database (NVD), which introduce latency between vulnerability disclosure and actionable risk assessment. This paper presents PoC2Defense, a multi-source vulnerability intelligence framework that integrates early community signals from platforms such as Bugtraq and CXSecurity with authoritative records from CVE/NVD, CISA Known Exploited Vulnerabilities (KEV) catalog data, and patch commit analysis to construct a unified benchmark dataset. The framework addresses three predictive tasks: (1) forecasting whether a newly disclosed vulnerability will be actively exploited in the wild (KEV prediction), (2) classifying vulnerability weakness types (CWE classification) from early textual descriptions before official NVD enrichment, and (3) computing enterprise-level priority rankings that account for exploit signal strength, chain potential, and patch availability. We curate a dataset of 67 real-world vulnerabilities spanning 2021-2024, covering web applications, authentication libraries, cloud infrastructure, platform vendors, and IoT systems. Each record is enriched with 38 features across five layers: source metadata, CVE attributes, exploit signals, patch intelligence, and chain relationships. Experimental evaluation using XGBoost, Random Forest, Gradient Boosting, Logistic Regression, and Ridge Regression demonstrates that early Bugtraq signals contribute measurably to prediction quality. The KEV predictor achieves an F1 score of 0.991 and AUC-ROC of 0.982, outperforming CVSS-only baselines $(\mathbf{F 1}=\mathbf{0. 9 1 8})$ and PoC-existence heuristics $(\mathbf{F 1}=\mathbf{0. 9 1 4})$. Ablation analysis confirms that removing Bugtraq signal features reduces KEV prediction F1 by 3.3 percentage points, validating the hypothesis that community-sourced early indicators carry predictive value beyond official metadata. The CWE classifier attains 0.767 weighted F1 using XGBoost, and the enterprise priority ranker reaches 0.725 weighted F1. All code, data, and trained models are released publicly to facilitate reproducibility and future research in vulnerability intelligence.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.