The global vulnerability reporting pipeline has hit a catastrophic bottleneck, not because of sophisticated zero-days, but due to an influx of automated garbage. Automated filters at the National Vulnerability Database (NVD) and the Cybersecurity and Infrastructure Security Agency (CISA) have proven shockingly easy to deceive, legitimizing a flood of hallucinated threats. According to JFrog Security Research, a GitHub repository under the handle 'programmervuln/cveadvisory' single-handedly injected over 50 fake advisories into the system, targeting SQLite and other critical software. JFrog researcher Afek Berger correctly identifies this as 'LLM slop'—content that mimics the formal structure of a technical report but lacks any basis in reality.
The Anatomy of Fabricated Vulnerabilities
Technical forensics by JFrog reveals a total decoupling between these AI-generated alerts and the actual source code they claim to exploit. Take CVE-2026-51302: it arrived with a terrifying 9.8 Critical severity score and cited a heap use-after-free error in a function named 'exprComputeOperands'. The problem? As Afek Berger noted, that function doesn't exist in SQLite version 3.41.0. When researchers tested the supposed Proof of Concept (PoC) in isolated Docker containers, the queries ran perfectly without a single crash. The entire 'vulnerability' was a creative writing exercise by a language model.
All advisories in this repo seem AI generated when testing them with Gptzero.
This pattern of hallucination isn't an isolated glitch; it is the new baseline for noise. CVE-2026-51303, another 'Critical' threat, even claimed a patch had been issued to fix the non-existent bug. A diff analysis proved the 'fix' was as imaginary as the flaw. Security teams are no longer just fighting hackers; they are being buried under 'slop'—structured technical fiction that triggers every automated alarm in the building.
The Financial Toll of Automated Trust
The economic fallout of this incompetence is measurable in wasted engineering hours. Red Hat initially swallowed the bait, assigning CVE-2026-51302 a perfect 10.0 score before backtracking to 7.6. In the enterprise world, a 'Critical' flag triggers mandatory patching cycles, audits, and emergency meetings. By automating the ingestion of these reports without manual verification, NVD and CISA have effectively outsourced the vetting process to the end-user. JFrog's data shows that while detection tools like Gptzero flagged almost every one of these 50+ CVEs as AI-generated, they still bypassed the defensive gates of international cybersecurity governance.
The cited code didn't even exist in those versions or referenced unrelated logic.
This failure erodes the foundational trust in Open Source reporting. The official SQLite advisory page remains clean, yet the public record is polluted with 'placeholder' metadata and contradictory findings. For CTOs and technical leads, the takeaway is clear: blind reliance on CVE databases is now a liability. Security strategy must pivot toward a verification model centered on official vendor advisories and vetted security labs. While the industry speculates about AI writing the 'perfect exploit,' the immediate reality is more pathetic: organizations are currently paying their highest-paid engineers to debug hallucinations because the regulators failed to install a basic spam filter.