Two supply chain threats: open LLMs with zero guardrails, and AI-hallucinated CVEs poisoning NVD

Two incidents this week converge on the same risk vector: AI systems inserting unverified information into critical security infrastructure—with no detection layer in place.

The Facts Two distinct signals converged this week. A SaferAI audit of GLM-5.2 (Zhipu) revealed offensive capabilities comparable to GPT-5.5 but a complete absence of content filters: 76% of tested vulnerabilities are reproducible, with zero refusals. Additionally, JFrog documented 55 CVEs submitted by a single account for SQLite—54 out of 55 are AI hallucinations, some of which passed NVD’s initial triage and ended up in reference databases.

Our Take These two incidents reveal that the security supply chain is not equipped to handle AI-driven vectors. Vulnerability scanners ingest NVD feeds without human validation—if fake CVEs slip through, they contaminate the security pipelines of thousands of companies. The GLM-5.2 case raises a different question: open-weight, auditable models expose the safety gaps that closed models conceal. The risk isn’t theoretical: both vectors are operational today.

To Watch The response from MITRE/NVD regarding the validation process for CVEs submitted via automated means, and regulatory pressures on open-weight model providers lacking guardrails.