threatdeck
Live 13:45 UTC ·

AI Stack Alert Critical CVE-2026-116692 For AppSec, AI Eng, SOC / IR, CISO AI-assisted

MLflow 3.12.0–3.16.x: unauthenticated RCE via third-party scorer deserialization (CVE-2026-116692)

MLflow 3.12.0 added third-party scorer deserialization. The server imports a request-controlled module path under four allowed package prefixes and instantiates a request-controlled class. A caller can create an experiment whose artifact location sits inside an allowed package directory, upload a Python module as a run artifact, and then trigger its import through the scorer invoke endpoint. Affected versions are mlflow >= 3.12.0rc0 and < 3.17.0, and the fix is in 3.17.0.

Why it matters

On default no-auth MLflow servers, a remote caller can run arbitrary Python code with the server process's privileges. The official image runs as uid 0, so an attacker could take over the host or container and reach model, experiment and artifact data. With basic auth enabled, the scorer endpoint does no resource-level permission check, so any user who can create experiments and upload artifacts may still reach it.

Do this

  1. Upgrade mlflow to 3.17.0 or later on all tracking servers.
  2. Until patched, do not expose MLflow servers to untrusted networks, and enable authentication and network access controls.
  3. Run the MLflow server as a non-root user and make the installed Python package tree read-only.
  4. Review experiments for artifact_location values that point into site-packages or mlflow/genai/scorers directories.
  5. Check server filesystems for unexpected run-ID directories under mlflow/genai/scorers/* and for unusual requests to /ajax-api/3.0/mlflow/scorer/invoke.

Exploit Clock

Exploitation unlikely (for now)

  • EPSS 0.70%

Signals

  • NoCISA KEV
  • 9.8CVSS
  • 0.70%EPSS
  • 45/100Priority
  • 90%Confidence

Evidence

“On the default no-auth server using the official MLflow v3.14.0 image, this produces unauthenticated arbitrary Python code execution with the privileges of the MLflow server process.”

GitHub Security Advisories (AI packages)

“A package-name prefix is not a provenance boundary. If the server process can write below an allowed package, an attacker-created descendant can satisfy the string-prefix check and become importable Python code.”

GitHub Security Advisories (AI packages)

“Affected: mlflow >= 3.12.0rc0, < 3.17.0 → fixed in 3.17.0.”

GitHub Security Advisories (AI packages)

Sources

Affected / related

MLflow, mlflow.genai.scorers, /ajax-api/3.0/mlflow/scorer/invoke

Published 2026-10-27 21:26 UTC · updated 2026-11-08 02:38 UTC · AI-assisted summary, checked against quoted sources · How we score

Following

Follow the vendors, products, AI tools, techniques and actors you care about. Matching stories, CVEs and exploited bugs are marked, a Following tab appears in the feed, and returning visits show what changed. Saved only in this browser.

Presets

Get updates

A personal RSS feed of new stories on everything you follow. Paste it into any feed reader, Slack or Teams RSS app.