JFrog Security Research investigated a set of SQLite vulnerability advisories published by a newly created GitHub account (programmervuln/cveadvisory-). NVD flagged them as critical and CISA's ADP agreed, but verification fell apart: the cited code did not exist in the named versions, PoC payloads triggered no crash, none of the CVEs appear on SQLite's official advisory page, and GPTZero testing indicated the texts were AI-generated.

Six CVEs were analyzed in detail. CVE-2026-51302 (9.8 Critical) claims a use-after-free via exprComputeOperands() — a function that did not exist in SQLite 3.41.0 and was only added in mid-2025; sqlite3ReleaseTempReg() merely recycles register indices, making a UAF impossible. CVE-2026-51303 (9.8) references a fix in 3.51.3, yet the diff between 3.51.2 and 3.51.3 shows no changes to src/expr.c, and its PoC is invalid SQL. CVE-2026-51300 (9.1) cites line numbers in expr.c that point to a comment and an unrelated allocation. CVE-2026-51297 (8.8) names jsonBlobEdit(), which did not exist in 3.41.0. CVE-2026-51296 (7.5) cites lines 3555/3575 of json.c in a file that is only 2706 lines long. CVE-2026-51304 (7.5) uses a function signature that does not exist, while the real code nulls the pointer immediately after deletion. All PoCs were executed in Docker containers under AddressSanitizer against official builds of versions 3.41.0, 3.51.2 and 3.51.3 — none triggered any memory error. Red Hat initially scored CVE-2026-51302 as 10.0 Critical, later downgrading it to 7.6 High.

JFrog attributes the incident to a broken pipeline: MITRE's submission form requires no identity verification, and since NIST paused deep NVD analysis in February 2024, plausible-sounding fake advisories can flow into GHSA, downstream databases and enterprise scanners without reproduction of any bug. A broader audit of 55 advisories from the same account found 54 completely fabricated and only one containing a real bug wrapped in unverified metadata.

The researchers list red flags: no mention on official maintainer security pages, no linked commits or pull requests, contradictory CPE metadata, and references to functions or line numbers that do not exist. They warn that fabricated critical CVEs waste triage and patching effort — and that AI agents used for automated remediation may attempt to patch non-existent code, making things worse. The findings were reported to GHSA, Red Hat and NVD.