Methodology
This is the reference set I use before I report an AI security incident. The public desk also polls a shorter RSS list. A name here is a source I read. It is not a claim that every one of them is in that automatic poll.
Advisory identifiers, severity scores, and the exploited-vulnerability catalog.
Primary notices from the companies that ship the software.
What OpenAI and Anthropic publish about attacks on or with their systems.
Research teams that document campaigns, tooling, and intrusion sets.
Disclosure programs aimed at model and application flaws.
National alerts and advisories.
Public catalogs of AI incidents and adversary behavior.
The tags on an incident use these two, and only when the report supports the category.
These are on the public desk feed. A rewrite of someone else's story does not count as a second outlet.
Cryptography and security papers, read as research rather than as a confirmed incident.
I do not label an incident critical unless one of these is true: a primary source confirms that severity, CISA lists it in the Known Exploited Vulnerabilities catalog, or at least two independent outlets report the incident. A CVSS of 9.0 or higher on the advisory itself counts as primary-source confirmation.
Two outlets means two separate newsrooms. A rewrite of the same story does not count. If none of those bars are met, the severity stays high, medium, low, or unknown, even when one headline uses the word critical.
When I have a vendor bulletin, a GitHub Security Advisory, or another primary document, that URL is the primary link and it is listed before news coverage. A news article stays on the row as the report I read. If I only have the news article, the primary link stays empty. I do not guess a vendor URL.
When a report is wrong, I update that incident in place. The id stays the same, so a citation to the row still resolves. Threat Stats shows a new last-updated time, and the incident feeds pick up the change. I do not delete a public row to hide a mistake.
AI Security Guru