A recent analysis by cybersecurity researchers has identified a batch of apparently fabricated vulnerabilities, many of which were assigned critical or high severity ratings, that entered the National Vulnerability Database (NVD) and other security advisories. These reports, which appear to be AI-generated, highlight significant weaknesses in the current vulnerability disclosure pipeline.
Software supply chain security firm JFrog reported last week that 54 alleged vulnerabilities originating from a new, obscure GitHub repository were largely bogus. Among these, six specific CVEs targeting the open-source SQLite database were found to be entirely unreproducible upon testing. These SQLite CVEs carried CVSS scores ranging from 7.5 to 9.8. One alleged use-after-free vulnerability, initially assigned a maximum CVSS score of 10.0 by Red Hat before being downgraded, referenced a function that does not exist in the specified SQLite version. Another use-after-free vulnerability, with a CVSS score of 9.1, cited unrelated source code lines, and its accompanying proof-of-concept executed a valid query without demonstrating any memory leaks or errors. The remaining four SQLite CVEs tested by JFrog were similarly found to be fake.
The GitHub repository in question also listed 49 other CVEs claiming to affect the open-source RAW image processing library libraw and the Arduino audio decoding library ESP32-audioI2S. While JFrog did not conduct extensive testing on these, the firm indicated that most are likely also fabricated, with one exception that contained a genuine bug but was wrapped in unverified CVE metadata.
The incident underscores a systemic issue with automated vulnerability ingestion. The process for assigning CVEs often operates on an honor system, where CVE Numbering Authorities (CNAs) trust requesters to have verified the information they provide, as CNAs are frequently not in a position to independently verify every report.
Historically, the US National Institute of Standards and Technology (NIST), which manages the NVD, provided a crucial backstop by manually reviewing and enriching CVE records. However, this process has been severely hampered since 2024 due to a surge in vulnerability submissions and operational challenges. By late 2024, NIST's backlog of unprocessed CVEs exceeded 17,000, growing to over 27,000 by the end of 2025, according to a May 2026 Department of Commerce Inspector General report. The report also criticized NIST for a "lack of strategic planning and decisive action" in addressing the backlog, leading to continued growth despite allocated funds. This means the pipeline currently lacks a mandatory checkpoint for independent reproduction of claimed vulnerabilities.
JFrog researchers noted that because no step in the current system explicitly requires a proof-of-concept or bug reproduction, a plausible-sounding but fake advisory can easily propagate through the pipeline, appearing in GitHub Security Advisories, downstream databases, and enterprise scanners. This situation can lead to security professionals wasting valuable time investigating non-existent threats.
To mitigate this risk, JFrog recommends several verification steps before acting on new CVEs. These include checking if the affected vendor has corroborated the issue (e.g., SQLite maintainers do not list these fake CVEs), looking for a lack of commit hashes or pull requests in reference fields, identifying suspicious metadata such as missing CPE product definitions, and verifying that code references actually match real functions or relate to the supposed issue.
JFrog reported its findings to the GitHub Security Advisory team, Red Hat, and NVD. Red Hat and NVD have reportedly flagged or removed the CVEs, though the GitHub repository itself remained active at the time of reporting.
The motivation behind these fake submissions is speculative, with possibilities including attempts to artificially boost research experience or to influence automated CVE identification tools. Regardless of the intent, security researchers anticipate an increase in such incidents, as generative AI has significantly reduced the effort required to produce plausible-looking advisories, while the effort to verify them remains substantial. This asymmetry presents a growing challenge for the cybersecurity industry.
