Research/Software
Security advisories that dependency scanners never see
Across 30 self-hosted AI/ML projects, 195 of 618 published security advisories (31.6%) never reached GitHub's global Advisory Database under their own ID. A pre-registered follow-up found the gap belongs to self-hosted applications, not to AI.
Oct 14, 2026 · 6 min read

When a maintainer publishes a security advisory on a GitHub repository, it protects users only if it also reaches the global GitHub Advisory Database, the source that Dependabot, npm audit, pip-audit and OSV-based scanners read. We counted how often that fails. Across 30 widely used self-hosted AI/ML repositories, 195 of 618 published repository advisories (31.6%) were not in the global database under their own ID, so those scanners could not flag them. The median missing advisory had been missing for 91.1 days.
We then asked whether this is a problem of AI tools. It is not. A pre-registered comparison with matched non-AI projects found no AI effect, and a second pre-registered census found a large, significant split between self-hosted applications and libraries: 53.0% of application advisories were invisible to package scanners, against 0 of 11 for libraries.
31.6%
of 618 AI/ML repository advisories not in the global database under their own ID
90
of those 195 had no global record under any ID
53.0% vs 0%
scanner-blind share, applications (482 of 909) vs libraries (0 of 11)
p = 0.0057
repository-level permutation test for the application-library gap
The question
A repository security advisory lives at a project's own security page. GitHub may review it and import it into the global Advisory Database, and only then do package scanners learn about it. The most detailed public model of that review pipeline collects repository advisories only through records already in the global database, so an advisory that was never imported cannot appear in it by construction. We found no published count of how many maintainer-published advisories never arrive.
What we did
On 14 October 2026 we listed every published advisory in 30 self-hosted AI/ML repositories and looked each one up in the global database by its own GHSA ID. Nineteen repositories had at least one advisory. For each missing advisory we checked whether it had any global presence under another ID, such as a thin NVD-sourced record for its CVE.
Two follow-ups were pre-registered: the hypothesis, sample, measure, test and decision rule were fixed before any data was collected. The first compared the 30 AI/ML repositories with 30 matched non-AI ones. The second compared 40 fresh self-hosted applications with the 40 most-depended-on libraries. In both, an advisory counts as scanner-blind when no OSV record, searched by its GHSA ID and every CVE alias, names a package. The unit of the significance test is the repository, with 10,000 label permutations.
What we found
Of the 195 missing advisories, 90 had no global record at all and 18 carried a CVE with only a thin record under a different ID; the other 87 had no CVE to cross-reference. The gap is not a queue. When GitHub does import a repository advisory, it does so quickly, with a median of 3.2 days from publication to review. Seven projects had every one of their advisories missing.
Share of each repository's advisories missing from the global database
The first follow-up tested whether AI tools are worse. They are not. Pooled, AI/ML advisories were less often scanner-blind than those of matched non-AI projects, and at the repository level the difference was not significant (p = 0.30). We also found, after the run, that the study script over-counted blind advisories by ignoring OSV records that GitHub's reviewers had fixed; corrected, the shares are 28.0% and 46.4%, and the conclusion does not change. Among the non-AI controls, a pattern stood out: 64% of self-hosted applications' advisories were missing from OSV, against 2% of libraries'.
Pre-registered comparisons, scanner-blind share
| Comparison | Group A | Group B | Repository-level p |
|---|---|---|---|
| AI/ML vs matched non-AI (17 Oct 2026) | AI/ML: 34.7% of 596 | Non-AI: 56.8% of 1,352 | 0.30 |
| Applications vs libraries (18 Oct 2026) | Applications: 53.0% (482 of 909) | Libraries: 0.0% (0 of 11) | 0.0057 |
The application-library split was then tested on new repositories, and it held: +53.0 percentage points, p = 0.0057, and still significant after dropping the most advisory-heavy repository in each group (p = 0.0199). Age explains part of the application figure but not all of it. Application advisories at least 90 days old were still blind 36.8% of the time.
Scanner-blind share of application advisories, by age at snapshot
What it means
The likely mechanism is simple. A library publishes to a package registry, which gives GitHub's reviewers a package to attach the advisory to. An application shipped only as a Docker image, a binary or source has no such package, so its advisory stays on the repository page where no package scanner looks. Within applications, those that also publish a registry package, such as Gogs or Payload, were mostly visible; Immich, Stirling-PDF, paperless-ngx and LibreChat were 100% blind. Even where these advisories carry a CVE, the OSV records we checked list only Git commit ranges, which no manifest scanner can match to a running container.
If you run self-hosted applications, a clean scanner report says little about them; watch their security pages directly. If you maintain one, check whether your published advisories resolve in the global database, and name a package when one exists. We publish a maintainer checklist and agent skills that run these checks for a single advisory or CVE.


