AI Sec Watch dataset

Programmatic access

No authentication is required, the limit is 60 requests per minute, and CORS is enabled. See the full API docs.

Quick start

import pandas as pd

# Load full dataset (CSV)
df = pd.read_csv("https://aisecwatch.com/api/issues/export")

# Or use the JSON API with filtering
import requests

resp = requests.get("https://aisecwatch.com/api/v1/issues", params={
    "severity": "critical",
    "limit": 100,
})
data = resp.json()["data"]

What's included

How it works

1. 1.

   Collect

   83 configured data sources fetched every six hours, including NVD, CISA KEV, GitHub Advisory, arXiv, and cybersecurity and AI news feeds. View all sources.

2. 2.

   Deduplicate

   Incoming items are deduplicated by URL and CVE ID to prevent double-counting.

3. 3.

   Classify

   Each item is processed by an LLM using a versioned prompt (currently v4) that assigns severity, issue type, attack type, affected vendors, and a confidence score (0 to 1).

4. 4.

   Enrich

   Vulnerabilities with CVE IDs are enriched with EPSS exploit probability scores (FIRST API), CISA Known Exploited Vulnerabilities catalog status, CWE-to-CAPEC attack pattern mapping, and CVSS vector parsing.

5. 5.

   Deliver

   Data is accessible via the web interface, public REST API (60 req/min, CORS enabled), STIX 2.1 threat intelligence feed, webhooks with HMAC-SHA256 signing, RSS feeds by category, and CSV export.

Enrichment details

EPSSExploit Prediction Scoring System

30-day exploit probability from FIRST. Scores range from 0 to 1, indicating how likely a vulnerability is to be exploited in the wild within the next 30 days.

CISA KEVKnown Exploited Vulnerabilities Catalog

Known exploitation status from CISA, including whether the vulnerability has been observed in active exploitation and whether it is associated with ransomware campaigns.

CAPECCommon Attack Pattern Enumeration and Classification

Attack pattern taxonomy mapping derived from CWE IDs. Maps weaknesses to known attack patterns to help defenders understand exploitation techniques.

CVSS v3.xCommon Vulnerability Scoring System

Vector string parsing for attack characteristics including attack vector, complexity, privileges required, user interaction, and impact metrics.

MITRE ATLASAdversarial Threat Landscape for AI Systems

AI/ML-specific attack taxonomy mapping. Attack types identified by the classifier are mapped to ATLAS technique IDs, enabling alignment with enterprise AI risk frameworks.

Versions

A release is a frozen copy with its own DOI, archived on Zenodo, and it does not change. Cite a release when an analysis must be repeatable. The downloads above are generated from the live database at the time of the request, so their record counts change after each fetch.

Latest release, any version

10.5281/zenodo.19519469

Classifier prompt, live data

v4

First archived release

2026-04-11

Live data last updated

Nov 4, 2026