About
About
This is a map of public AI security failures: vulnerabilities in the software that runs AI, and documented attacks on AI systems. Each record is filed under a class and a technique so you can see how a kind of failure actually happens.
Where the records come from
- MITRE ATLAS case studies: 77 records, used under Apache 2.0.
- GitHub Advisory Database: 2,259 records, used under CC BY 4.0.
- HackerOne public reports: 16 records, linked only, with the class guessed from the title.
- Exploited status comes from the CISA Known Exploited Vulnerabilities catalogue, and likelihood scores from EPSS.
- One-line summaries and class filings for advisories are written by a model that read each advisory. They are ours, not the source's, and can be wrong.
- Payload passages come from PayloadsAllTheThings by Swissky and contributors, used under the MIT licence.
How records are grouped
Classes follow the OWASP Top 10 for LLM Applications and MITRE ATLAS. Advisories are filed by a model that reads each one, so some are in the wrong place, and the record page says how each was filed. Case studies are grouped by the ATLAS techniques their authors assigned.
Data last built 2026-11-17. Records belong to their sources and authors. Use what you find here only on systems you are authorised to test.