A batch of SQLite CVEs rated critical and high appeared in the NVD — the U.S. government's reference vulnerability database — with CISA-supplied enrichment, only to turn out to be technically bogus. JFrog researchers, who reported the case on July 30, reviewed the cited code and tried to reproduce the exploits, only to find nonexistent functions, references to unrelated sources, and proof-of-concept payloads that simply did not work. The verdict was categorical: fabricated vulnerabilities, very likely generated by AI. And the case was not isolated — the same researchers flagged dozens of suspicious 'slop' findings, some already carrying official CVE identifiers and critical ratings, including a batch in which 54 of 55 submissions made by a single account were fake. The incident draws attention because SQLite is one of the most ubiquitous pieces of software on the planet, embedded in browsers, phones, and operating systems — and any noise made in its name ripples across the entire ecosystem.
The CVE pipeline is the backbone of modern security. When a flaw is discovered, it gets a CVE identifier and enters the NVD with metadata, references, and a CVSS score. Teams around the world consult these databases to decide what to patch first. The problem is that the submission process through MITRE's public form requires no real identity verification: virtually anyone can propose a description and suggest a CVSS rating. AI automation has removed the last cost that kept the door closed — the difficulty of writing a plausible report. Now language models produce convincing descriptions, point to code that does not contain the flaw, and invent functions that never existed. Without manual validation by a human who actually reads the code, the garbage gets in.
There are telltale signs of this 'slop,' such as a lack of commit hash or pull request in the reference fields and suspicious metadata with missing CPE product definitions. But relying on every scanner to catch those patterns is fragile. The scale is precisely what makes the situation dangerous: enterprise scanners and vulnerability-management platforms pull the NVD daily and turn every entry into a remediation task. A single false positive labeled 'critical' can trigger overnight alerts, mobilize on-call staff, and generate hours of rework. The cumulative effect is security fatigue: the more noise, the less trust — and that trust is what sustains the whole patch-prioritization flow. When even the 'trusted' data contains garbage, teams chase flaws that do not exist or, worse, ignore real warnings because they have learned to doubt the database. The integrity of the CVE ecosystem underpins compliance and audit decisions; if it is called into question, everything loses credibility.
The phenomenon extends beyond vulnerability databases. Apple, overwhelmed by a surge in AI-generated security reports, has begun capping submissions to its bug bounty program and imposed a 30-day waiting period. In the meantime, a legitimate macOS flaw worth up to $200,000 ended up blocked and went unreported. As generative tools flood every reporting channel, the tension between openness and noise grows harder to manage. The open question remains: how do we tell, at scale, what is a real finding from a well-written hallucination — without strangling precisely the researchers who deliver value?
Sources: The Register, JFrog Security Research, Cybernews, 9to5Mac
✓ Independent sources cross-checked and verified before publishing
