ns0bj/cve-threat-intelligence-hub

cve · threat-intelligence · vulnerabilities · cisa-kev · epss · metasploit · nuclei · exploit · vector-search · rag

---

🛡️ Global CVE & Threat Intelligence Dataset with 384-dim Embeddings (1999–2026)

The most comprehensive, unified, and actionable open-source cybersecurity dataset indexing all 400,000+ published CVEs enriched with multi-source threat intelligence, a ready-to-query 1.4 GB SQLite database, and 384-dimensional dense semantic embeddings generated via sentence-transformers/all-MiniLM-L6-v2.

Official GitHub Repository: nsozturk/cve-threat-intelligence-hub
Live Documentation Portal: https://nsozturk.github.io/cve-threat-intelligence-hub/

---

📊 Dataset Summary

This repository hosts all artifacts needed to run 100% offline threat intelligence, RAG security copilots, and vulnerability research:

Source / Telemetry	Count	Description
Authoritative CVE Records	400,776	Complete historical archive from MITRE / CVE.org (1999–2026)
FIRST EPSS Probabilities	381,682	Machine-learned 30-day exploit prediction probabilities & percentiles
CISA KEV (Known Exploited)	1,730	Vulnerabilities confirmed actively exploited in the wild
Rapid7 Metasploit Modules	3,521	Weaponized exploit modules mapped directly to CVEs
ProjectDiscovery Nuclei	4,509	Automated scanning and verification templates
GitHub PoC Repositories	25,524	Public GitHub exploit repositories with stars & fork metrics
Exploit-DB Archive	46,689	Verified exploits, shellcodes, and academic security papers
Debian Security Tracker	71,525	Operating system package patch statuses across Debian releases
MITRE CWE Definitions	969	Root-cause weakness classifications (SQLi, RCE, IDOR, etc.)
Open Source Advisories	699,127	PyPI, npm, Go, Maven, Rust, Linux kernel advisories

---

📦 Repository Files & Formats

This Hugging Face repository provides multiple formats for every workflow:

1. cve.db (1.4 GB SQLite Database):
   • Production-ready SQLite 3.33+ database in WAL mode.
   • Includes all tables (cves, cisa_kev, epss_scores, metasploit_modules, nuclei_templates, github_pocs, exploits, debian_security, cwes) and a standalone FTS5 full-text search virtual table.
2. cve_embeddings.npy (293.5 MB NumPy Matrix):
   • Dense unit-normalized embeddings for all 400,776 CVEs in float16 precision (shape=(400776, 384)).
   • Enables cosine similarity search across all CVEs in ~280ms on Apple Silicon Metal (mps) or CPU.
3. cve_ids.json (6.7 MB JSON):
   • Array of 400,776 CVE IDs precisely mapped 1:1 to the rows of cve_embeddings.npy.
4. cves_metadata.parquet (47.7 MB Parquet):
   • Fast, lightweight columnar metadata table without vectors for DuckDB, Pandas, and Polars.
5. data/train-*.parquet (8 chunks, ~600 MB):
   • Native Hugging Face Datasets format containing the full metadata plus the embedding vector column as a PyArrow fixed-size list (list[float32, 384]).

---

🗄️ Parquet Schema

Column	Type	Description
cve_id	string	Unique CVE Identifier (e.g. CVE-2024-3094)
title	string	Vulnerability title or headline
description	string	Full technical vulnerability description
cvss_score	float32	CVSS base severity score (0.0 to 10.0)
cvss_version	string	CVSS version (e.g. 3.1, 4.0, 2.0)
cvss_vector	string	CVSS metric vector string
epss_score	float32	EPSS exploitation probability in the next 30 days (0.0 to 1.0)
epss_percentile	float32	EPSS percentile ranking relative to all scored CVEs
is_cisa_kev	int8	1 if listed on CISA Known Exploited Vulnerabilities catalog, else 0
has_exploit	int8	1 if any verified exploit code or PoC exists, else 0
has_metasploit	int8	1 if a weaponized Rapid7 Metasploit module exists
has_nuclei	int8	1 if a ProjectDiscovery Nuclei detection/exploit template exists
has_poc	int8	1 if a public GitHub PoC repository exists
date_published	string	ISO timestamp of initial publication
date_updated	string	ISO timestamp of last update
assigner	string	CNA organization (e.g. cisco, microsoft, redhat, apache)
vendor_product	string	Target vendor and affected product name
metasploit_module	string	Path to official Metasploit module (e.g. exploit/linux/http/...)
nuclei_template	string	Path to Nuclei YAML template (e.g. http/cves/2024/...)
top_poc_url	string	URL of top-starred GitHub PoC repository
top_poc_stars	int32	Number of GitHub stars on the top PoC repository
embedding	list[float32, 384]	Dense 384-dimensional unit-normalized semantic embedding vector

---

🚀 Quickstart & Usage Examples

1. Load with Hugging Face datasets

from datasets import load_dataset

# Load dataset (all 400k CVEs with embeddings)
ds = load_dataset("ns0bj/cve-threat-intelligence-hub", split="train")
print(f"Loaded {len(ds):,} CVE records")

sample = ds[0]
print(sample["cve_id"], sample["title"], sample["cvss_score"], sample["epss_score"])

---

2. High-Speed Sub-Second Semantic Search (cve_embeddings.npy)

Search by describing an exploit scenario without needing exact keywords:

import json
import numpy as np
from huggingface_hub import hf_hub_download
from sentence_transformers import SentenceTransformer

# 1. Download embedding matrix and IDs
repo = "ns0bj/cve-threat-intelligence-hub"
npy_path = hf_hub_download(repo_id=repo, repo_type="dataset", filename="cve_embeddings.npy")
ids_path = hf_hub_download(repo_id=repo, repo_type="dataset", filename="cve_ids.json")

embeddings = np.load(npy_path).astype(np.float32)  # shape: (400776, 384)
with open(ids_path) as f:
    cve_ids = json.load(f)

# 2. Encode query with sentence-transformers
model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
query = "remote code execution in fortinet ssl-vpn authentication bypass"
query_vec = model.encode([query], normalize_embeddings=True)[0]

# 3. Compute cosine similarities in ~280ms
scores = np.dot(embeddings, query_vec)
top_indices = np.argsort(-scores)[:5]

for idx in top_indices:
    print(f"[{scores[idx]*100:.1f}% Match] {cve_ids[idx]}")

---

3. Direct SQLite Database Queries (cve.db)

Download and query the complete relational database directly without building from scratch:

import sqlite3
from huggingface_hub import hf_hub_download

db_path = hf_hub_download(
    repo_id="ns0bj/cve-threat-intelligence-hub",
    repo_type="dataset",
    filename="cve.db"
)

con = sqlite3.connect(db_path)
cur = con.cursor()

# Find actively exploited vulnerabilities with high EPSS probability
query = """
SELECT id, title, cvss_score, epss_score, has_metasploit, has_nuclei
FROM cves
WHERE is_cisa_kev = 1 AND epss_score > 0.85
ORDER BY epss_score DESC
LIMIT 10
"""
for row in cur.execute(query):
    print(row)

---

4. Serverless Analytics with DuckDB

Run fast SQL queries directly on remote or local Parquet files without full RAM loading:

import duckdb

con = duckdb.connect()
df = con.execute("""
    SELECT cve_id, title, cvss_score, epss_score, top_poc_url, top_poc_stars
    FROM 'hf://datasets/ns0bj/cve-threat-intelligence-hub/cves_metadata.parquet'
    WHERE is_cisa_kev = 1 AND epss_score > 0.8
    ORDER BY epss_score DESC
    LIMIT 10
""").df()

print(df)

---

📈 Use Cases

• RAG for Security Copilots: Provide ground-truth vulnerability information and threat telemetry to local LLMs (Claude Code, Cursor, Codex, Antigravity) with sub-second retrieval.
• Prioritization & Vulnerability Management: Combine EPSS probability (>80%), CISA KEV status, and CVSS scores to triage real-world risk rather than theoretical severity.
• Threat Hunting & Exploit Verification: Instantly verify if weaponized Rapid7 Metasploit modules, Nuclei templates, or GitHub PoCs exist for any discovered vulnerability.

---

⚖️ License & Attribution

This dataset is released under the MIT License.

• CVE Records: CVE Project (MITRE / CVE.org)
• EPSS Scores: FIRST.org / Cyentia Institute
• CISA KEV: US Cybersecurity and Infrastructure Security Agency
• Exploit-DB: OffSec Exploit Database
• Nuclei: ProjectDiscovery
• Debian Security: Debian Security Tracker
• OSV: Google Open Source Vulnerabilities