Over the past few days, a newly created GitHub repository published batch advisories for SQLite vulnerabilities that were quickly marked critical by the National Vulnerability Database (NVD) and approved by CISA’s Authorized Data Publisher program. However, when JFrog security researchers investigated, they found the claims did not hold up to scrutiny.
According to JFrog’s analysis, cited code did not exist in the targeted SQLite versions, proof-of-concept payloads failed to trigger crashes, and none of the CVEs appear on SQLite’s official advisory page. Testing the advisories with Gptzero, an AI-detection tool, suggested all advisories were AI-generated. When combined into a single file, the advisories triggered AI-generated content warnings.
JFrog tested six specific CVEs with these findings:
CVE-2026-51302 claimed a use-after-free in exprComputeOperands(), but that function did not exist in SQLite 3.41.0—it was added in mid-2025. Additionally, sqlite3ReleaseTempReg() only recycles register indices and does not perform heap deallocation, making the claimed vulnerability impossible.
CVE-2026-51303 alleged ExprListDelete() fails to clear back-references. JFrog found no evidence of such pointers in relevant structures, and a diff between versions 3.51.2 and 3.51.3 showed no changes to the relevant code file.
CVE-2026-51300 cited specific line numbers in expr.c (1012 and 1026) as evidence of a use-after-free bug, but those lines contain a comment and a memory allocation call unrelated to the claimed vulnerability.
CVE-2026-51297 referenced jsonBlobEdit(), which was not present in the targeted version 3.41.0.
CVE-2026-51296 cited line numbers 3555 and 3575 of json.c, but version 3.41.0’s json.c file is only 2706 lines long.
CVE-2026-51304 claimed a use-after-free but the single-argument function signature reported does not exist—the actual function requires a database pointer parameter.
JFrog’s broader audit of 55 advisories from the same GitHub account found 54 were completely fabricated, with only one containing a real bug.
The vulnerability submission process through MITRE lacks identity verification, allowing anyone to submit advisory descriptions and propose CVSS scores. NVD, historically a validation layer, effectively paused deep analysis in February 2024 due to a surge in reports, according to the research. With no requirement for proof-of-concept or bug reproduction in the current system, plausible-sounding fake advisories can reach GHSA, downstream databases, and enterprise scanners.
Key facts
- 54 of 55 CVEs audited from the repository were completely fabricated
- None of the fake CVEs appear on SQLite’s official advisory page
- JFrog found cited code did not exist in targeted versions or proof-of-concept payloads failed to execute
- The submission process lacks identity verification and NVD paused deep analysis in February 2024
- Red Hat initially scored one CVE (CVE-2026-51302) as 10.0 Critical but later downgraded it to 7.6 High