54 AI-Fabricated Vulnerability Reports Reach the National Vulnerability Database

SQLite, the libraw RAW image processing library, and the ESP32-audioI2S Arduino audio decoder, plus the CVE assignment pipeline itself

JFrog examined 55 vulnerability reports posted from a newly created GitHub account over a matter of days and found 54 of them fabricated. None of the six SQLite reports, which carried CVSS scores ranging from 9.8 down to 7.5, described a reproducible vulnerability. Run through an AI-detection checker, the advisories came back as likely machine-generated. The fakes did not stay on GitHub: they propagated into GitHub Security Advisories and then into the National Vulnerability Database, the authoritative US government record, where NVD marked several critical and a CISA team enriched them. MITRE subsequently rejected the entire repository, Red Hat lowered one score, and NVD flagged or removed the records, though GitHub had not removed the repository at the time of reporting. The practical hazard runs downstream: security teams increasingly point AI coding agents at fabricated advisories, and an agent handed a nonexistent CVE will hunt for a vulnerable function nobody wrote and modify working code to patch a problem that does not exist

Anyone can submit a CVE through MITRE's public form and the assigning bodies largely take the submitter's word. Generative AI has driven the cost of producing a plausible-looking advisory to nearly zero while the cost of verifying one is unchanged, against an NVD backlog that exceeded 27,000 unprocessed records by the end of 2025