Introduction
In the high-stakes environment of cybersecurity, timing is the ultimate defensive advantage. When a new vulnerability (CVE) is discovered, it is first announced by MITRE. However, there is a significant lag before the U.S. National Institute of Standards and Technology (NIST) releases an official Common Vulnerability Scoring System (CVSS) report—averaging 132 days.
During this four-month "blind spot," Enterprise Security Officers (ESOs) are forced to prioritize patches without a clear understanding of a bug's severity, while attackers may already be developing exploits. VEST (Vulnerability Exploit Scoring & Timing) is a predictive system designed to close this gap by providing early, automated predictions of exploit timing and severity using real-time social media data.
System Architecture
VEST operates as an integrated pipeline that bridges raw data collection with actionable security intelligence. The system's architecture is designed to handle the high velocity of social media data while maintaining a structured database for historical analysis.
As illustrated in the architecture, the system consists of three main layers:
- Vulnerability Database: A back-end "Data Crawler" continuously monitors MITRE, NIST, and Twitter, storing data in a SQLite database organized into node and edge tables.
- Predictive Engines: These engines process the raw data to predict exploit probability, timing, and CVSS severity scores.
- User Interface: A web-based dashboard that allows users to query vulnerabilities, visualize relationships via 3D force graphs, and view detailed prediction reports.
Methodology
VEST utilizes two primary predictive components to transform social media "chatter" into technical metrics:
Exploit Timing (The CAT Graph)
To predict if and when an exploit will appear, the authors developed a CVE-Author-Tweet (CAT) graph. This model captures the complex relationships between vulnerabilities, the authors discussing them, and the tweets themselves. By using recursive equations to calculate "popularity" scores for these entities, the system generates features that can predict exploit emergence with high precision.
Severity Estimation (GCN + Attention)
To estimate the CVSS score (0–10) and its eight security attributes, VEST uses an attention-based feature embedding to extract technical nuances from tweet content. These embeddings are fed into a Graph Convolutional Network (GCN). The GCN allows the system to exploit intrinsic correlations between different vulnerabilities—for instance, recognizing that similar buffer overflow bugs in different software packages likely share severity characteristics.
Key Results and Performance
The researchers evaluated VEST using 23 months of CVE data, establishing that the security community's collective discussion on Twitter acts as a real-time proxy for the technical interest and exploitability of a bug.
Key performance metrics include:
- Early Warning: Just 3 days of Twitter activity following a CVE announcement is sufficient to generate a highly accurate severity estimate.
- Exploit Timing: VEST predicts the date a real-world exploit emerges with an accuracy of $\pm 11.90$ days.
- Severity Accuracy: The system predicts CVSS severity scores with a Mean Absolute Percentage Error (MAPE) of 0.18.
Limitations and Future Work
While VEST offers a massive head start over official NIST data, it faces certain challenges. The system is currently dependent on Twitter; if the cybersecurity community shifts to other platforms, the models would require recalibration. Additionally, future work must address "adversarial noise," where attackers might intentionally hype minor bugs to distract defenders. Finally, while a $\pm 12$-day window is a vast improvement over 132 days, narrowing this window further remains a priority for zero-day defense.
Summary updated