54 of 55 SQLite CVEs were AI-fabricated; one rated 10.0 Critical
JFrog found a repo of hallucinated SQLite flaws that sailed into NVD and Red Hat feeds, poisoning the scanners and patch-bots your pipeline trusts.
Copy markdown
One phantom flaw, a perfect 10.0
JFrog dissected 55 SQLite advisories from one GitHub repo and found 54 were fabricated: cited code that never existed, proof-of-concepts that never crashed. Red Hat still stamped CVE-2026-51302 a 10.0 Critical before quietly cutting it to 7.6.
Why LLM slop still earns a CVE ID
The assignment pipeline verifies no identity and demands no working repro, and NIST's manual validation has been backlogged since Feb 2024. An AI-written report can mint a Critical CVE that your dependency scanner surfaces as a five-alarm emergency.
Your patch-bot may chase ghosts
Blanket 'patch every Critical' rules, driven by insurance, compliance and audits, now point remediation at code that isn't there, and AI agents can 'fix' nonexistent bugs and add real risk. Meanwhile genuine reports go unacknowledged while maintainers drown in the noise.
What to do before you patch
Treat Critical CVEs as claims, not facts. SQLite's own page says most CVEs against it aren't real vulnerabilities and that you should not assume a CVE about SQLite is authoritative. Cross-check sqlite.org/cves.html and require a working repro before you auto-remediate.