Critical CVE Issued for a Hallucinated SQLite Vulnerability

A new GitHub repository published over 50 vulnerability advisories, including some for SQLite, and some of the SQLite CVEs were reflected in NVD and CISA security metadata. However, verification by JFrog researchers found that the vulnerability claims and PoCs did not match the actual code.

The verification results are as follows.

• The functions or code cited in the advisories either did not exist in the relevant version or pointed to logic unrelated to the claimed vulnerability.
• Running the provided SQL PoCs against official SQLite builds with AddressSanitizer (ASan) produced no memory errors or crashes.
• The CVEs in question were not listed on SQLite's official security advisory page.
• Throughout the advisories, errors commonly seen in AI-generated documents were identified, such as nonexistent fixes, incorrect function arguments, and placeholder metadata.

JFrog directly inspected the 3.41.0, 3.51.2, and 3.51.3 tags of the official SQLite repository, built the release versions in an isolated Docker environment, and attempted to reproduce the PoCs. The CVEs analyzed included the exprComputeOperands() UAF in CVE-2026-51302, the ExprListDelete() UAF in CVE-2026-51303, and a UAF related to JSON processing functions.

Notably, CVE-2026-51302 was initially rated CVSS 10.0 Critical but was later downgraded to 7.6 High. This case shows that vulnerability advisories that appear to be LLM-generated can enter CVE and security databases before verification, highlighting the need for code- and PoC-based validation of automated vulnerability classification systems.