AI slop pollutes the CVE pipeline with fake vulns - The…
AI-generated fake vulnerabilities targeting SQLite and other libraries highlight critical weaknesses in the CVE pipeline and automated vulnerability…
- Incident date
- Aug 2026
- Target
- SQLite
In August 2026, security researchers identified a batch of critical and high-rated CVEs targeting SQLite that were discovered to be entirely fabricated. These synthetic advisories, which appeared in the National Vulnerability Database (NVD) with CISA-supplied enrichment, exposed significant vulnerabilities in the automated systems used to track software security flaws.
What happened
Software supply chain firm JFrog reported that six SQLite vulnerabilities, carrying CVSS scores between 7.5 and 9.8, were generated by an obscure GitHub repository. Testing revealed the reports were likely AI-generated, as they cited non-existent functions or irrelevant source code lines. In one instance, a reported use-after-free vulnerability failed to trigger any memory issues during proof-of-concept testing. Beyond SQLite, the same repository contained 49 additional fake CVEs targeting libraries like libraw and ESP32-audioI2S.
The incident highlights a systemic failure in the CVE ingestion process, which operates largely on an honor system where reporters are trusted to verify their own findings. This is compounded by the NVD’s struggle with a massive backlog of unprocessed records, which grew to over 27,000 by late 2025. Because the current pipeline lacks a mandatory requirement for independent bug reproduction or proof-of-concept validation, plausible-sounding AI-generated advisories can bypass existing security filters and propagate into enterprise scanners and downstream databases.
Experts suggest that the ease of generating fake advisories via generative AI creates a dangerous asymmetry, as the effort to produce a report is near zero, while the effort to verify and reproduce one remains significant. To mitigate these risks, defenders are encouraged to verify that vulnerabilities are corroborated by the software vendor, check for legitimate commit hashes or pull requests in reference fields, and scrutinize metadata for inconsistencies. Following the disclosure, MITRE rejected the repository's submissions, and entities like Red Hat and the NVD moved to flag or remove the bogus entries. However, the incident serves as a warning that automated vulnerability ingestion remains highly susceptible to pollution in the AI era.