Someone posted 55 vulnerabilities in days. 54 were AI-invented. One still scored a perfect 10.

The software industry runs on a shared list of known bugs. Someone is now seeding that list with breakage that no one ever found.

Security firm JFrog found six critical SQLite flaws that were pure fiction, it reported. The advisories claimed severe memory bugs with scores up to 9.8. But the functions they blamed did not exist in the versions cited, and the proof-of-concept code ran without a crash. Fed through the detector GPTZero, the write-ups came back as machine-generated. They were AI slop.

And not just six. One new GitHub account posted 55 vulnerability reports in a matter of days, and JFrog found 54 fabrications. Only one held a real bug, wrapped in bogus metadata.

The rest invented holes in SQLite, the image library libraw and an Arduino audio decoder. Red Hat briefly rated one SQLite entry a perfect 10.0, then lowered it to 7.6. SQLite added the function it blamed in 2025, a year after the version it supposedly broke.

How fiction became official

The fakes did not stay on GitHub. They reached the National Vulnerability Database, the US government’s authoritative record. There the NVD marked them critical, and a CISA team enriched them. Once a vulnerability lands there, scanners and defenders worldwide treat it as real.

It works on trust. Anyone can submit a CVE through MITRE’s public form, and the bodies that assign them largely take the submitter’s word. “The CNA is often not in a position of being able to verify the report themselves,” noted Oracle engineer Alan Coopersmith.

NIST used to be the backstop. Its experts reviewed each record by hand. Then a flood of submissions forced them to pause deep analysis in early 2024. The backlog passed 27,000 unprocessed reports by the end of 2025. A federal watchdog said the agency wasted money failing to clear it. No step in the pipeline now requires anyone to reproduce a bug.

The machines fool the machines

The real danger is what reads these entries next. Security teams increasingly point AI coding agents at their code to fix reported flaws. Hand one a fabricated CVE and it may hunt for a vulnerable function nobody ever wrote. Then it changes working code to patch a problem that does not exist.

The loop barely has a human in it. One AI writes a plausible advisory, a thinly staffed pipeline waves it through, and another AI tries to fix it. The same fluency without substance already dogs AI-written code.

JFrog’s Afek Berger framed it as an asymmetry. Generating a convincing fake now costs almost nothing, he said. Verifying one costs exactly what it always did: read the source, build the version, run the proof of concept.

Even well-resourced defenders cannot check every report by hand. CISA and NIST are too far behind to help. The honour system never anticipated machines that can lie at scale.