Security firm JFrog — known for its software‑supply‑chain tools — has identified a striking surge of false vulnerability disclosures that slipped through the public CVE pipeline. In a matter of days, a newly created GitHub account published 55 separate reports. JFrog’s analysis confirmed that 54 of those entries were pure fiction, their source code never containing the alleged flaws.

The fabricated reports covered a range of software components. Most claimed memory‑corruption bugs in the SQLite database engine, but the cited functions did not exist in any released version. Others targeted the image‑processing library libraw and an Arduino‑based audio decoder, again describing functions that were never part of the code base. One entry even received a perfect CVSS 10.0 rating from Red Hat before being revised to 7.6 after the error was spotted.

From GitHub to the National Vulnerability Database

After the initial posting, the bogus CVEs migrated beyond GitHub. The U.S. National Vulnerability Database (NVD), the federal repository that many organizations rely on for patch prioritization, accepted the entries and marked them as critical. A CISA team later enriched the records, adding metadata that security scanners around the world automatically trust.

The pipeline that governs CVE assignment is built on a trust model. Submitters fill out a public form managed by MITRE, and the designated CNA (CVE Numbering Authority) typically accepts the information without independent verification. As Oracle engineer Alan Coopersmith noted, “The CNA is often not in a position of being able to verify the report themselves.” Historically, NIST staff reviewed each record manually, but a flood of submissions forced the agency to pause deep analysis in early 2024. By the end of 2025, a backlog of more than 27,000 unprocessed reports lingered, according to a federal watchdog.

Because no step in the current process requires a reproducible proof‑of‑concept, the fabricated entries slipped through unchecked. The false CVEs now sit in the NVD, where vulnerability scanners treat them as real threats, prompting organizations to waste time investigating non‑existent bugs.

AI‑generated fakes and the automation loop

JFrog ran the suspect advisories through GPTZero, a detector that flags machine‑generated text. The tool labeled the write‑ups as AI‑originated. JFrog’s Afek Berger described the phenomenon as an asymmetry: “Generating a convincing fake now costs almost nothing; verifying one costs exactly what it always did—reading the source, building the version, running the proof‑of‑concept.”

The danger escalates when automated remediation tools, such as AI‑driven code‑fixers, consume these bogus CVEs. An AI assistant handed a fabricated vulnerability may scour a code base for a function that never existed, then apply a patch that could break legitimate functionality. The loop—AI writes a fake advisory, a thinly staffed pipeline passes it, another AI attempts a fix—creates a feedback cycle with little human oversight.

Even well‑funded security teams cannot manually verify every incoming report. CISA and NIST, already stretched thin, lack the capacity to act as a reliable backstop. The incident underscores a systemic weakness: the honor‑based submission system never anticipated machines capable of mass‑scale deception.

JFrog’s discovery serves as a warning that the vulnerability‑management ecosystem must adapt. Without stronger verification steps, AI‑generated fakes could continue to erode confidence in the very databases that organizations depend on to safeguard their software.

Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.