updated
Editorial policy
How the register is researched, what each label means, and how corrections work.
Sources we use
For definitions and controls we use the framework bodies themselves: the OWASP GenAI Security Project's Top 10 lists, MITRE ATLAS, NIST's AI 600-1 profile, and the Model Context Protocol specification. For cases we prefer, in order: the affected organization's own disclosure, a CVE or government advisory, then the researchers' write-up. Security vendors publish much of the research in this field and sell products related to it; source lists note that interest.
How claims are checked
Every case summary is written in our own words from the original disclosure, which we read in full unless the source list says otherwise. When a primary source would not load, the case says which secondary source we relied on. For academic preprints we note when only the abstract was read. Numbers appear only when the source gives them, with the conditions attached (for example, a cost estimate that the source calls a worst case).
Labels
- Incident and research demo follow MITRE ATLAS's Incident and Exercise labels. An incident happened in the wild; a research demo showed an attack working, often in a controlled setting.
- Source types (standard, government guidance, incident disclosure, vendor research, independent research, preprint, news) appear on every source list.
- Framework mappings say whether they come from OWASP's own appendix or are our choice. The table of which risks apply to which uses is our judgment.
- Vendor questions and anything marked "editorial guidance" are our suggestions, not framework text.
What we don't publish
We don't publish working attack payloads, product rankings, or claims of testing we haven't done. A risk with no case listed means we haven't added a verified one, not that it is rare.
Updates and corrections
The register is checked against its sources on the date shown at the bottom of every page. We re-check when OWASP or MITRE ATLAS publish a new release, and at least once a quarter. Guides list substantive changes in an update history; typo fixes don't change the date. Report an error to info@aisecurityrisks.com and we will correct it and note the change.