VulGD: A LLM-Powered Dynamic Open-Access Vulnerability Graph Database

Luat Do, Jiao Yin, Jinli Cao, Hua WangarXiv 2026 · DOI

TL;DR

This paper introduces VulLink, a dynamic, open-access vulnerability graph database that integrates data from multiple public repositories. It addresses the challenge of fragmented cybersecurity data by providing a continuously updated graph structure, pre-computed semantic embeddings of vulnerability descriptions, and an accessible API for downstream data mining tasks.

What Problem Does This Paper Address?

The ever-increasing volume of software vulnerabilities presents a significant challenge for cybersecurity professionals and researchers. While numerous public repositories like NVD, CVE, and ExploitDB offer valuable information, this data is often siloed, record-centric, and difficult to correlate. This fragmentation makes it challenging to capture complex relationships between vulnerabilities, weaknesses, exploits, affected products, and vendors, hindering effective threat intelligence analysis and data mining. Existing graph-based resources often suffer from being static, offline, or difficult to access, limiting their utility for dynamic and reproducible cybersecurity analytics.

Methods

VulLink tackles this problem through a multi-faceted approach:

  • Dynamic Multi-Source ETL Pipeline: An automated Extract-Transform-Load (ETL) pipeline is designed to continuously ingest data from various public vulnerability sources. This pipeline normalizes disparate data formats, standardizes identifiers, and constructs typed entities and explicit relationships, transforming isolated records into a unified graph structure. The pipeline supports incremental updates, ensuring the graph remains current without requiring full rebuilds.
  • Vulnerability Graph Database: The core of VulLink is a Neo4j-based property graph database. This structure models entities like Vulnerabilities, Exploits, Weaknesses, Products, and Vendors as nodes, connected by relationships such as EXPLOITS, AFFECTS, and BELONGS_TO. This allows for complex, multi-hop queries that reveal intricate connections.
  • Pre-computed Semantic Features: To enrich the graph data, VulLink generates embeddings of vulnerability descriptions using pre-trained language models (e.g., SecBERT, all-mpnet-base-v2). These embeddings capture semantic nuances not always present in structured fields, providing valuable features for downstream machine learning tasks. The system supports various embedding dimensions and models, allowing users to choose based on their needs.
  • Open Access Layer: VulLink provides both an interactive Web interface for visual exploration and a public API for programmatic access. This layer enables users to query the graph, download customized subgraphs, and retrieve pre-computed embeddings, facilitating both exploratory analysis and automated data mining workflows.

Datasets

VulLink integrates data from the following public vulnerability repositories:

  • National Vulnerability Database (NVD)
  • Common Vulnerabilities and Exposures (CVE)
  • Common Weakness Enumeration (CWE)
  • Exploit Database (ExploitDB)
  • CVE Details

The paper does not introduce new benchmark datasets but rather leverages existing public sources to construct its comprehensive graph database.

Key Insights and Ideas

  • The Power of Dynamic Graph Integration: The paper highlights that static snapshots of vulnerability data quickly become obsolete. A continuously updated graph database is crucial for real-time cybersecurity analytics.
  • Bridging the Gap Between Siloed Data and Graph Mining: VulLink demonstrates a practical solution for converting fragmented, record-centric vulnerability information into a queryable graph structure, making it accessible for advanced data mining.
  • Semantic Embeddings Enhance Graph Context: Combining structured graph data with semantic embeddings from vulnerability descriptions significantly improves the performance of downstream tasks like exploitability prediction.
  • Open Access is Key for Reproducibility and Adoption: Providing a public API and Web interface lowers the barrier to entry for researchers and practitioners, fostering wider adoption and enabling reproducible research.
  • Trade-offs in Embedding Dimensionality: The study shows a clear trade-off between embedding dimensionality, storage costs, and retrieval time, guiding users in selecting appropriate dimensions for their specific use cases.

Key Results

  • The deployed VulLink graph contains over 343,598 vulnerability nodes and 1,660,599 relationships, representing a significant scale of real-world vulnerability data.
  • In an exploitability prediction case study, a graph-augmented learning model (HGNN) significantly outperformed a baseline MLP model that only used tabular attributes.
  • The HGNN model achieved substantial improvements in PR-AUC (up to +257.8%), Macro-F1, and Precision@100, particularly when using only description embeddings, demonstrating the value of graph propagation.
  • The best-performing model (HGNN with combined features) achieved 44.89% PR-AUC, 64.62% Macro-F1, and 63.10% Precision@100, showcasing the utility of VulLink's features for practical vulnerability triage.

Limitations and Future Work

  • Scalability: The authors acknowledge that scalability under continuous graph growth is a challenge that will require further attention.
  • Integration of More Sources: Future work aims to integrate additional vulnerability and threat intelligence sources to further enrich the graph.
  • Extended Graph Mining Utilities: The team plans to develop more graph mining utilities, including those for dependency-aware risk analysis and graph-evidence-grounded patch prioritization.
  • Optimal Feature Fusion: The study noted that simple concatenation of embeddings and tabular features didn't always yield the best results, suggesting that more sophisticated fusion strategies might be beneficial for downstream tasks.

Summary updated

BibTeX
@misc{do2026vulgd,
  author = {Luat Do and Jiao Yin and Jinli Cao and Hua Wang},
  title = {{VulGD: A LLM-Powered Dynamic Open-Access Vulnerability Graph Database}},
  year = {2026},
  eprint = {2604.06967},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.06967},
}
Figure