What Happened to Hallucinated Critical CVEs for SQLite?

The cybersecurity community faced a significant disruption in late July 2026 when a series of critical Common Vulnerabilities and Exposures (CVEs) related to the widely used SQLite database engine were published and rapidly disseminated. These advisories, originating from a newly created GitHub repository, claimed to detail severe memory corruption and use-after-free vulnerabilities in various SQLite versions. Initially, these CVEs gained traction, being flagged as critical by the National Vulnerability Database (NVD) and CISA's Automated Indicator Sharing (ADP) program.

The rapid acceptance of these seemingly critical vulnerabilities raised alarms, prompting deeper investigation. JFrog Security Research quickly stepped in to verify the claims, only to discover that the reported vulnerabilities were entirely fabricated. Their analysis revealed that the code paths and functions cited in the advisories did not exist in the specified SQLite versions, and proof-of-concept payloads failed to trigger any crashes or demonstrate the described flaws. JFrog's researchers termed these as 'LLM slop,' suggesting they were likely generated by large language models without factual basis.

This incident was a direct consequence of systemic issues within the vulnerability reporting ecosystem. A critical factor was the significant backlog at NIST's NVD, which, since February 2024, had paused its deep analysis and enrichment efforts for incoming CVEs due to an overwhelming surge in reports. This meant that many vulnerabilities were being processed with less manual scrutiny, creating an opening for unverified or fabricated entries to gain official recognition. The CVE submission process itself, lacking robust identity verification, further contributed to the ease with which such 'slop CVEs' could be introduced.

The consequences of these hallucinated CVEs were immediate and far-reaching. Security teams, relying on automated feeds and the perceived authority of NVD and CISA, wasted valuable resources investigating non-existent threats, attempting to patch non-existent code, and generating internal tickets. This not only diverted attention from genuine vulnerabilities but also risked eroding trust in the CVE system itself. The event served as a stark reminder of the dangers of AI hallucinations in high-stakes environments like cybersecurity, where incorrect information can lead to misallocated resources, false alarms, and potentially overlooked real threats.

As of August 3, 2026, the cybersecurity community is grappling with the implications. There's an increased emphasis on the need for human-in-the-loop verification for AI-generated security intelligence and a renewed call for improvements in the CVE vetting process. While AI continues to show promise in discovering real vulnerabilities (as demonstrated by Google's Big Sleep finding a legitimate SQLite zero-day in late 2024), this incident underscores the critical importance of distinguishing between AI-assisted discovery and AI-generated fabrication. SQLite itself continues to release updates addressing real, often AI-discovered, bugs, while also maintaining its own authoritative list of vulnerabilities to counter misinformation.