31.6% of security advisories for self-hosted AI tools never reach scanner databases

The gap arises because many maintainers publish advisories on GitHub without the ecosystem metadata or CVE assignment that feeds databases such as OSV and the GitHub Advisory Database's reviewed set. Tools like Dependabot, Snyk-style SCA and container scanners only match what is in those feeds. Nearly a third of the advisories are effectively invisible to automated pipelines.

This matters because self-hosted AI stacks are proliferating inside enterprises, often deployed by data teams outside normal AppSec review. Inference servers like vLLM and low-code agent builders like Flowise and LangFlow frequently expose network ports and hold API keys.

The finding fits a week heavy on AI-security themes. AWS pitched continuous penetration testing with Continuum, and a CSO Online piece arguing that labs skip security basics drew 837 HN points. The finding also ties to concerns about agents leaking credentials.

What to do: teams should subscribe directly to project security advisories rather than relying only on scanners. Watch whether GitHub or OSV respond with ingestion fixes.