AI-Generated Fake Vulnerabilities Pollute Security Databases
Fake security advisories created by AI threaten the CVE ecosystem.
1Discovery of Fake Vulnerabilities
JFrog's research revealed that six SQLite vulnerabilities published on GitHub were completely fake. These reports had CVSS scores ranging from 9.8 to 7.5 but none described an actual security flaw. Believed to be AI-generated, the reports relied on functions that didn't exist in the affected SQLite version. For example, one alleged use-after-free vulnerability was initially given a 10.0 CVSS score by Red Hat before being downgraded.
2Other Fake Vulnerabilities in GitHub Repo
The suspicious GitHub repository contained 49 additional vulnerability claims for libraw and ESP32-audioI2S libraries. While JFrog didn't extensively test these, it determined all but one were fake. MITRE rejected all vulnerabilities from this repo, exposing weaknesses in the security advisory system. Oracle Solaris engineer Alan Coopersmith noted that MITRE and other CVE assigners typically don't verify submissions.
3NVD Verification Process Issues
The National Vulnerability Database (NVD), managed by NIST, slowed its verification process in 2024 due to increased submissions and operational challenges. By late 2024, over 17,000 unprocessed CVEs had accumulated. Despite plans to clear the backlog, it grew to 27,000 by 2025's end. A Department of Commerce Inspector General report found NIST inefficiently used funds allocated for backlog clearance due to lack of strategic planning.
4Lack of Mandatory Verification
The current CVE system lacks mandatory steps to independently verify every claimed vulnerability. JFrog stated this allows fake reports to reach GitHub Security Advisories, databases, and enterprise scanners. This incident highlights systemic issues with automated vulnerability ingestion, forcing security professionals to waste time on fake reports instead of real threats.
5Recommended Precautions Against Fake CVEs
JFrog recommended several checks before acting on new CVEs: vendor corroboration (e.g., SQLite maintainers not listing fake CVEs) and checking for missing commit hashes or suspicious metadata (like absent CPE product definitions). Code references not matching real functions or pointing to irrelevant code sections also indicate likely AI-generated fake reports.
6AI's Role in Vulnerability Reporting
JFrog researcher Afek Berger noted generative AI has nearly eliminated the effort needed to create plausible-looking advisories while verification still requires significant work. This asymmetry means even well-resourced defenders can't manually validate every report. Berger emphasized this as an industry-wide challenge in the AI era.
7GitHub's Response and Ongoing Risks
JFrog reported findings to GitHub Security Advisory, Red Hat, and NVD. Most flagged or removed the CVEs, but GitHub's repo remains accessible. JFrog speculates fake reports might aim to boost research credentials or influence automated CVE tools. While motivations remain unclear, the incident demonstrates how AI can exploit vulnerability reporting systems.
8A Warning for the Security Industry
These 54 fake CVEs represent just one example of AI's potential misuse in security reporting. JFrog predicts similar incidents will increase. The industry needs stronger verification mechanisms to combat fake reports. This event serves as a wake-up call about AI's threats to cybersecurity infrastructure and the need for preparedness.