AILab project
AI Security Tracker: methodology
01Cite
How to cite this tracker.
8020 Lab. AI Security Tracker. https://8020lab.com/ai-incidents/ (accessed [date]).
For one entry, cite its own page: https://8020lab.com/ai-incidents/entries/<id>/. The full data set is at /ai-incidents/data/entries.json; new entries are in the feed.
02Scope
What counts as an AI security incident.
Security incidents and vulnerabilities in AI products: language-model services, agents, coding assistants, MCP servers and clients, AI browsers and the frameworks under them, published since 1 January 2025.
- Automated rows: NVD exact-phrase searches, kept only when the description names the AI layer (phrases such as “prompt injection”, “model context protocol” or “language model”; acronyms such as LLM, MCP and RAG; or a tracked package). A sentence that only credits an AI tool for finding a bug does not qualify.
- GitHub Advisory Database and OSV records for a fixed list of AI packages.
- Curated incidents without a CVE, each checked against a primary source: the vendor, a regulator, a court or the researcher’s own write-up.
- Duplicates merge by CVE and GHSA identifier; a curated incident absorbs the CVEs it links; the CISA KEV catalog marks what is exploited.
03Sources
Where the rows come from.
| Source | Fetched | Records |
|---|---|---|
| NVD CVE API 2.0 | 2026-11-02 | 1615 exact-phrase matches, 1583 AI-layer |
| GitHub Advisory Database | 2026-11-02 | 1840 advisories, 30 packages |
| OSV | 2026-11-02 | 1503 records, 30 packages |
| CISA KEV | 2026-11-02 | catalog 2026.10.30, 1766 entries |
| Curated incidents | 2026-11-05 | 22 entries, each checked against its primary source |
04Labels
What each label means.
- Research disclosed
- Published by the researcher; no vendor statement found.
- Vendor confirmed
- The vendor or maintainer acknowledged the issue.
- Fixed
- A fix, patched version or mitigation is published.
- Alleged
- Reported by secondary sources only; the primary source could not be confirmed.
- Disputed
- The vendor or another party contests the report.
- Exploited
- Listed in the CISA Known Exploited Vulnerabilities catalog, or exploitation in the wild is stated by the primary source.
Severity is the CVSS base score when published: NVD first, then the GitHub or OSV record. The word breach is used only when a vendor, regulator or court confirmed unauthorized access.
05At the last build
2,002 entries, updated 2026-11-05.
- CVE records
- 1,966
- GitHub advisories
- 14
- Curated incidents
- 22
- Exploited
- 17
06Limits
What this tracker is not.
- Not complete: a term filter misses some AI-layer issues and admits some that are not.
- Not a ranking: counts per vendor reflect how much a vendor publishes, not how safe its products are.
- Automated rows repeat the database text; read the primary source before acting.
- 283 pages with little text of their own (entries with a summary under 120 characters, and vendor pages with a single entry) stay readable but ask search engines not to index them.