JFrog: 54 of 55 advisories from one account were fabricated

01What happened

JFrog Security Research published the first part of a series on AI and CVEs on October 27, 2026. Tamir Friedman, Afek Berger and Aviv Engelberg set out five pressures they attribute to AI: a lower barrier to filing reports, convincing fakes, severity inflation, faster exploitation and a growing patch queue. The figures they cite are listed below.

02How it works

Fabrication is the clearest case. JFrog reviewed 55 advisories that a newly created GitHub account published in one batch and found 54 completely fabricated, with details the authors say appear to have been hallucinated by a language model: one advisory cited line 3555 of a file that has 2706 lines. The reports looked polished and authoritative, carried scores up to 9.8, and CISA's ADP program agreed with those scores.

03Who is affected

Teams that triage by CVSS base score are the ones this reaches, along with anyone relying on NVD enrichment. NIST moved CVEs published before March 1, 2026 into a not-scheduled queue when it began prioritising KEV entries, federal software and critical software in April 2026, so many records never receive NVD analysis.

04What to do

Treat a CVSS base score as a starting point rather than an urgency signal, and check whether a report comes with a working proof of concept before acting on it. JFrog argues that exploitability in your own configuration should drive the queue.

Questions people ask

How did JFrog find that 54 of 55 advisories were fabricated?

JFrog Security Research checked each advisory that a newly created GitHub account published in one batch against the SQLite source. The authors report that 54 of the 55 were completely fabricated, citing functions that do not exist and impossible line numbers, including a reference to line 3555 in a file of 2706 lines. The reports carried scores as high as 9.8.

Should I still use CVSS base scores to prioritise patching?

JFrog argues that a CVSS base score alone should not drive urgency, because scores built on worst-case assumptions overstate risk. The authors point to 27 curl CVEs that maintainers rated Low to Medium and that CISA escalated to High or Critical, and found some Critical ratings came from client-side bugs classified as network-based. Exploitability in your own configuration is the better signal.

Why are many CVEs missing NVD analysis?

NIST changed NVD operations in April 2026 to handle record CVE growth, prioritising entries in CISA's KEV catalog, software used in the federal government and critical software under Executive Order 14028. Records published before March 1, 2026 moved to a category NIST labels lowest priority and not scheduled for immediate enrichment, so their details may stay unanalysed unless someone asks for them.

Sources 3 links, all fetched

1. 1jfrog.comJFrog Analysis
2. 2nist.govNIST NVD Update
3. 3depthfirst.comFFmpeg Zero-Days