AI-Generated Fake CVEs Pollute Vulnerability Databases
A wave of fabricated security advisories, likely created by AI, has infiltrated vulnerability databases, exposing critical flaws in the CVE submission and validation process.
The cybersecurity ecosystem is facing a new threat as artificial intelligence becomes a tool for generating entirely fake vulnerabilities. Security researchers at JFrog have uncovered a batch of six supposed CVEs for SQLite, along with others for libraw and ESP32-audioI2S, that were found to be completely fabricated. These bogus advisories had not only received high CVSS scores but also CISA enrichment, despite describing non-existent functions or unrelated code, highlighting systemic issues within the CVE pipeline.
These fabricated advisories, originating from an obscure GitHub repository, were initially assigned high CVSS scores, with one even receiving a maximum 10.0 score from Red Hat before being adjusted. Upon investigation, JFrog found that none of the six SQLite reports described a reproducible vulnerability. For instance, one alleged use-after-free flaw was linked to a function that did not exist in the targeted SQLite version, while another cited source lines unrelated to the purported issue. The accompanying proof-of-concept for this second flaw simply executed a valid query without any errors, further indicating its inauthenticity.
The problematic GitHub repository contained a total of 55 CVEs, with 49 of them targeting open-source libraries like libraw and ESP32-audioI2S. While JFrog did not conduct exhaustive testing on all of these, they indicated that the majority were also fake, with one exception that contained a genuine bug but was bundled with unverified CVE metadata. MITRE has since reportedly rejected the entire repository's submissions, but the incident serves as a stark warning about the integrity of vulnerability reporting.
This situation underscores a critical weakness in the CVE assignment process: a reliance on the 'honor system.' As noted by Oracle Solaris engineer Alan Coopersmith, many CVE Numbering Authorities (CNAs) trust requesters to provide verified information, as they are often not in a position to independently reproduce every reported vulnerability. This lack of mandatory verification allows plausible-sounding but false advisories to bypass checks and enter widely used databases.
The problem is exacerbated by the National Institute of Standards and Technology's (NIST) National Vulnerability Database (NVD). Historically, NVD provided a crucial backstop through manual review and enrichment. However, a surge in submissions and operational challenges led to a significant backlog, growing to over 27,000 unprocessed CVEs by the end of 2025. A Department of Commerce Inspector General report criticized NIST's handling of this backlog, citing a "lack of strategic planning and decisive action."
The absence of a mandatory checkpoint requiring independent reproduction of claimed vulnerabilities means that fake advisories can easily slip through the cracks. JFrog stated that this incident demonstrates a "systemic issue with automated vulnerability ingestion." This pollution of vulnerability databases wastes valuable time and resources for security professionals who must now sift through potentially fabricated threats to identify genuine risks.
To mitigate this risk, JFrog recommends several checks for defenders. These include verifying if the vendor has corroborated the issue, checking for commit hashes or pull requests in reference fields, and scrutinizing suspicious metadata like missing CPE product definitions. If code references do not align with actual functions or the supposed issue, it is likely an AI hallucination.
JFrog speculates that the motivation behind these fake CVEs could range from inflating research credentials to manipulating automated vulnerability detection tools. Regardless of intent, the ease with which generative AI can produce convincing-looking advisories, contrasted with the significant effort required for verification, creates an asymmetry that challenges the entire cybersecurity industry. This incident is likely just the beginning of a trend that will require significant adjustments to vulnerability management and validation processes.