
Detection-only triage scanner for MLflow scorer RCE CVE-2026-116692: version check via the /version endpoint. No payloads, no scorer invocation.
mlflow-scorer-scannerDetection-only triage scanner for MLflow third-party scorer deserialization RCE, CVE-2026-116692. It asks a tracking server one safe question, its version, and grades it. It never exploits anything.
The flawDetail | Value |
|---|---|
CVE | CVE-2026-116692 |
Advisory | GHSA-26p8-2jq9-3vq9 |
Bug | Deserialization of untrusted data (CWE-94), third-party scorer handling in mlflow/genai/scorers/base.py |
Impact | Unauthenticated remote code execution on a default tracking server, root in the official image |
Severity | 9.8 critical |
MLflow 3.12.0 introduced registration and deserialization support for third-party scorers. The vulnerable code accepts a request-selected module path beneath one of four allowed package prefixes instead of requiring an exact allowed module, then imports that module and constructs the class during Scorer.model_validate_json(). An attacker who can place a Python module under a writable allowed package, via the normal artifact endpoints, and then submit it through /ajax-api/3.0/mlflow/scorer/invoke, gets code execution in the server process. This scanner never does any of that: no scorer request, no experiment, no run, no artifact upload.
Two source disagreements, and what this scanner follows. One write-up frames the range as 3.12.0 through 3.16.x; the advisory text itself, quoted in a second write-up, starts at 3.12.0rc0 and ends before 3.17.0. This scanner follows the advisory text, so 3.12.0rc0 grades VULNERABLE. On authentication, the advisory text says auth-disabled deployments permit the attack without credentials, while basic-auth deployments require a principal able to create experiments and runs and upload artifacts. A version in range on an auth-enabled server is therefore still a finding, but a weaker one than the same version exposed without auth.
Fixed versionsBranch | Status |
|---|---|
MLflow < 3.12.0rc0 | NOT_AFFECTED, scorer support did not exist yet |
MLflow 3.12.0rc0 to before 3.17.0 | VULNERABLE |
MLflow >= 3.17.0 | FIXED |
Updating to 3.17.0 is the fix. Until then, do not expose a tracking server without authentication, and make sure no scorer package path is writable by the server process.
What it checksGET <base>/version, the endpoint MLflow itself documents for checking a tracking server version. That is the only request it makes.JSON wrapper like {"version": "3.14.0"} is accepted; anything else unparseable grades UNKNOWN.What it never doesNo payloads. No scorer invocations. No experiment, run, or artifact creation. No upload of any file, malicious or otherwise. It reads one public version string and stops.
GradingFinding | Severity |
|---|---|
Version 3.12.0rc0 to before 3.17.0 | VULNERABLE |
Version 3.17.0 or later | FIXED |
Version before 3.12.0rc0 | NOT_AFFECTED |
Version hidden behind auth, 404, or unparseable | UNKNOWN, not clean |
Target unreachable or not http(s) | ERROR, not clean |
Usage# Triage one tracking server (default MLflow port is 5000)
python3 mlflow_scorer_scanner.py http://localhost:5000
# Machine-readable output
python3 mlflow_scorer_scanner.py https://mlflow.example.com --json
# Tests (mock servers only, no real MLflow needed)
python3 -m unittest discover -s tests -v
Exit codes: 0 = fixed, not affected, or unknown, 2 = vulnerable finding, 1 = scan error.
Honest caveatsCVE at time of writing, and it was not in CISA KEV. That is a statement about what was public on 2026-10-28, not a safety claim.Nuclei template or widely-available public scanner specifically for CVE-2026-116692 was found at time of writing. Generic MLflow version and unauthenticated-access checks exist, but they do not grade this scorer range./version is blocked or auth-gated, this scanner says UNKNOWN, not clean. Confirm the version another way before closing the finding.ReferencesGHSA-26p8-2jq9-3vq9 (MLflow advisory, affected range, fixed in 3.17.0)CVE-2026-116692 record (published 2026-10-27, CVSS 9.8, CWE-94)MLflow tracking server docs (the /version endpoint)Thanks toMLflow maintainers for the advisory and the 3.17.0 fixMLflow tracking-server exposure researchDisclaimerBuilt for defensive triage and authorized testing only. Scan only systems you own or are authorized to test. The authors are not responsible for misuse.
Research by
Vineeth Kumar(foxcornlab). Found a bug or want a feature, open anissue.