AI-generated noise is inundating the CVE pipeline with counterfeit vulnerabilities.

      The software industry relies on a shared database of known bugs. Currently, that database is being populated with fabricated issues that no one has ever detected. Security company JFrog reported discovering six critical non-existent vulnerabilities in SQLite. The advisories falsely claimed serious memory issues with severity scores as high as 9.8. However, the functions referenced did not exist in the versions mentioned, and the proof-of-concept code executed without causing any crashes. Analysis with the detector GPTZero revealed that the reports were generated by a machine, essentially being AI-generated nonsense.

      This was not limited to six cases. A new GitHub account submitted 55 vulnerability reports in just a few days, and JFrog identified 54 of them as fabrications. Only one contained an actual bug, obscured within false metadata. The rest fabricated vulnerabilities in SQLite, the image library libraw, and an Arduino audio decoder. Red Hat temporarily rated one of the SQLite vulnerabilities a perfect 10.0 before downgrading it to 7.6. SQLite later introduced the function erroneously blamed for an issue in 2025, a year after the version it allegedly compromised.

      How these fabrications gained legitimacy is concerning. The fake reports didn’t remain confined to GitHub; they made their way to the National Vulnerability Database (NVD), the official record maintained by the US government, which marked them as critical. A team from CISA enriched these reports, and once a vulnerability is listed there, security scanners and defenders worldwide treat it as legitimate.

      The system operates on trust. Anyone can submit a CVE through MITRE’s public form, and the organizations responsible for assigning them generally rely on the submitter's credibility. Oracle engineer Alan Coopersmith noted that “the CNA is often not in a position to verify the report themselves.”

      NIST used to act as a safeguard, with experts manually reviewing each record. However, a deluge of submissions forced them to pause thorough evaluations in early 2024, resulting in a backlog of over 27,000 unprocessed reports by the end of 2025. A federal oversight agency criticized the organization for mismanaging resources without addressing the backlog. Currently, there is no requirement in the process for anyone to reproduce a bug.

      The real threat lies in the systems that will read these entries next. Security teams are progressively using AI coding tools to address reported flaws. If one of these tools encounters a fabricated CVE, it may search for a non-existent vulnerable function and modify functioning code to remedy a problem that doesn’t actually exist.

      The process has become less human-driven. One AI generates a convincing advisory, a minimally staffed pipeline assesses it, and another AI attempts to resolve it. This same lack of substance already characterizes code produced by AI. JFrog's Afek Berger emphasized this imbalance, noting that creating a credible fake now requires minimal resources, whereas verifying a claim still demands the same effort: reviewing the source, building the version, and executing the proof of concept.

      Even organizations with sufficient resources cannot manually verify every report. CISA and NIST are currently too overwhelmed to provide assistance. The reliance on an honor system did not anticipate the emergence of machines capable of lying on a large scale.