Fifty-four fabricated CVEs, and a pipeline that requires no proof of concept

A new GitHub account published a batch of SQLite advisories. NVD flagged them critical, CISA's ADP agreed, and Red Hat initially scored CVE-2026-51302 at 10.0 before downgrading to 7.6. JFrog checked out the official tags, built SQLite in clean Docker containers and fed every proof of concept verbatim into the binaries under AddressSanitizer. None crashed. The findings are specific: exprComputeOperands() did not exist in the claimed version 3.41.0, one advisory cites lines 3555 and 3575 of a json.c that is 2,706 lines long, and the fix claimed for CVE-2026-51303 is contradicted by a diff touching src/expr.c not at all. Across the 55 advisories from that account, "54 were completely fabricated, while one contained a real bug wrapped in unverified CVE metadata." The cause given is procedural: no identity check on MITRE's form, NIST's February 2024 pause on deep NVD analysis, and "no step in today's system actually requires a proof-of-concept or bug reproduction." JFrog sells Xray, an automated vulnerability-ingestion product, so it has an interest here. In the Hacker News thread, at 208 points and 62 comments at 09:14 EDT, bluGill reports the opposite trend: "A year ago most LLM found issues were false positives... but that has changed and now LLMs are finding real issues in code."

Why it matters: JFrog describes the consequence for agents directly: "An AI agent that encounters a fabricated CVE may attempt to locate the vulnerable function, generate a patch, or recommend changes based on code that does not even exist." Input generated, verification skipped, remediation generated, and nothing in the chain executes the artifact. The post's four red flags are the mechanisable defence: no vendor corroboration, no linked commit, contradictory CPE metadata, code references that do not resolve. If your pipeline opens tickets off CVSS scores, wiring those four in is a day's work.