Skip to content
James Mullen·Indexed by this site · Originally publishedDate not given· Event →· Its place in the scan record →

Edgescan study: CVE status is insufficient to measure AI false reports

AI summary

Edgescan researcher James Mullen published an empirical study concluding that CVE status data cannot measure AI-generated false vulnerability reports. The study is based on a cvelistV5 snapshot from October 29, 2026, classifies all 5,115 CVE rejection records from January 2024 to October 2026 (including 1,058 withdrawn records that had been published), and examines 49 reports flagged by curl maintainers as AI slop and 12 in-depth cases. The results show that none of the 5,115 rejection reasons mention AI, and at least 431 published records list an AI system as the finder or tool; in the curl window, 21 of 47 reports established AI involvement, and 9 cited code that does not exist or exists only in other versions; the most common failure was not crossing a security boundary. The study was completed and drafted by Anthropic's Claude under researcher guidance, and the data and code are publicly available and reproducible with the research package.

Read originaledgescan.com