Tuskira AdvisoryReported, not on KEVLast updatedNovember 11, 2026

CVE-2026-116692

MLflow remote code execution through third-party scorer plugins, plus a related file-read flaw (CVE-2026-116690)

Latest change
Previdian reports 403 exploitation attempts against CVE-2026-116692 from 26 IPs, October 28 to November 11. MLflow and CISA haven't confirmed it. Not on CISA KEV.

Reachable if an MLflow server on 3.12 to 3.16 can be reached over the network, authentication is off or many users can create experiments, and it can write to its own Python package directory

Affected
MLflow 3.12.0rc0 to 3.16.x, plus 3.10.0 to 3.16.x for the file-read flaw CVE-2026-116690
Exploitation
Reported, not on KEV. Previdian reports 403 exploitation attempts against its sensors from 26 IPs (Oct 28 to Nov 11), and also reports attempts on CVE-2026-116690. MLflow and CISA haven't confirmed exploitation, and no compromise has been reported.
Fix
MLflow 3.17.0, with a coordinated mlflow db upgrade on SQL-backed servers
Can't patch today
Run MLflow as non-root with a read-only package directory

CVE-2026-116692 is a critical remote code execution flaw in MLflow 3.12.0rc0 through 3.16.x. You're exposed if an affected MLflow tracking server is reachable over the network, authentication is off (the default) or many users can create experiments, and the server process can write into its own Python package directory. MLflow 3.17.0 fixes it, along with the related file-read flaw CVE-2026-116690 (7.5), which affects 3.10.0 through 3.16.x.

Previdian, a company that runs internet sensors, reports 403 exploitation attempts against CVE-2026-116692 from 26 IP addresses between October 28 and November 11, 2026. MLflow and CISA haven't confirmed exploitation, no successful compromise has been reported, and neither CVE is on CISA's Known Exploited Vulnerabilities catalog. A deliberately harmless public proof of concept is attached to the GitHub advisory. This advisory doesn't cover CVE-2026-64849, a different MLflow flaw (SSRF) that has been on CISA KEV since August 19, 2026; don't confuse the two.

Is it reachable for you?

How CVE-2026-116692 works. MLflow 3.12.0 added support for third-party scorers (evaluation plugins) from four allowed packages, ragas, deepeval, trulens and phoenix, and checked only that a requested module name started with one of those package names. Per the advisory, a caller can create an experiment whose artifact location points inside one of those package folders, create a run, upload a Python file as a run artifact, then call /ajax-api/3.0/mlflow/scorer/invoke. MLflow imports the uploaded file and runs its code inside the server process before checking anything else. The advisory reproduced this on the official ghcr.io/mlflow/mlflow:v3.14.0 image, which runs as root and can write to its own installed packages.

How CVE-2026-116690 works. A prompt-tagged model version can be created with a relative path such as ../../../tmp as its source, and /api/2.0/mlflow/model-versions/get-artifact then serves files from that location with the server's permissions. The advisory lists /etc/passwd, MLflow's SQLite databases (including the basic-auth user database with the admin password hash) and shell history as confirmed read targets.

Who needs credentials. On a server without authentication, the mlflow server default, neither flaw needs credentials. With basic auth on, CVE-2026-116692 needs an account that can create an experiment and run and upload artifacts, and CVE-2026-116690 needs an account that can create a registered model. MLflow's docs list Create Experiment and Create Registered Model as requiring no specific permission, so any signed-in user may qualify.

Checks, in order. Run these only against MLflow servers your organization owns.

  1. Find MLflow servers and their version. On each host or image, run mlflow --version or pip show mlflow. 3.12.0rc0 up to (not including) 3.17.0 is affected by CVE-2026-116692; 3.10.0 up to 3.17.0 by CVE-2026-116690. Search container registries and Helm values for ghcr.io/mlflow/mlflow tags in those ranges. From the network side, curl -s http://<host>:5000/version returns the server version and /health confirms an MLflow server is answering; these routes come from MLflow's source, and neither request changes anything.
  2. Check network exposure. The default binds port 5000 to 127.0.0.1, but containers and many deployments start with --host 0.0.0.0. Check whether the port is reachable from the internet, user networks or other workloads, and check reverse proxies, ingress controllers and load balancers that publish MLflow on 80 or 443. A --static-prefix moves every path under a prefix.
  3. Check whether authentication is on. Look at the start command or container args for --app-name basic-auth or a custom app. If it's missing, anyone who can reach the port can use both flaws. A harmless test from an untrusted network position: if curl -s -o /dev/null -w "%{http_code}" http://<host>:5000/api/2.0/mlflow/experiments/search?max_results=1 returns 200 instead of 401, the API is open.
  4. Check package-directory permissions (CVE-2026-116692). The chain needs a file-backed artifact location that resolves inside an installed scorer package, so the service user must be able to write there. Run python -c "import mlflow, os; print(os.path.dirname(mlflow.__file__))" as the service user and check whether it can write to mlflow/genai/scorers/ and its ragas, deepeval, trulens and phoenix subfolders. Check whether the process runs as root and whether the container root filesystem is read-only. Per the advisory, a read-only filesystem or a package directory owned by a different user blocks this exact chain.
  5. Check who can create things (authenticated servers). List users who can create experiments, runs and artifacts, and registered models. Review default_permission in the basic-auth configuration and any role grants. Each of those identities is a path in until you upgrade.
  6. Check what the server can read (CVE-2026-116690). The basic_auth.db user database, tracking databases, cloud credentials, environment files and mounted secrets. That list sets how urgent the fix is per server.
What Tuskira checks for this flaw in your environment
MLflow 3.12 to 3.16 servers reachable from user networks or the internet with authentication off
A package directory the MLflow server process can write to
Across every system you run.
See how Tuskira finds exposed systems

Compensating controls (test them, don't assume them)

These reduce exposure while you schedule the 3.17.0 upgrade. Test each one against your own server from the network positions that matter.

  • Restrict network access. Allow the MLflow port only from the hosts and networks that need it, and take any internet-facing server off the internet first.
  • Turn on authentication. Start with --app-name basic-auth, set MLFLOW_FLASK_SERVER_SECRET_KEY and a strong MLFLOW_AUTH_ADMIN_PASSWORD, and tighten default_permission. This narrows exposure from anyone to any user who can create experiments or models; it doesn't close it.
  • Block the two endpoints at the proxy or WAF. Deny any path ending in /mlflow/scorer/invoke and any path ending in model-versions/get-artifact (MLflow serves it under /api/2.0/mlflow/, /ajax-api/2.0/mlflow/ and the bare route, per MLflow's source). Match after any static prefix. Expect side effects: blocking scorer/invoke stops on-demand GenAI scoring, and blocking get-artifact stops model version artifact downloads through the server.
  • Run as non-root with a read-only package directory. Set a non-root runAsUser, a read-only root filesystem, and a site-packages owner different from the service user. Per the advisory this blocks the exact CVE-2026-116692 chain. It does nothing for CVE-2026-116690.
  • Watch the process. EDR or process monitoring on the MLflow server for spawned shells or unexpected child processes, and for new .py files under site-packages/mlflow/genai/scorers/.
  • Rotate what the server could read. If an affected server was reachable without authentication, plan to rotate basic-auth passwords and any cloud or database keys on disk or in its environment.

What to look for in logs

  • Calls to scorer/invoke, especially from sources that don't normally use GenAI evaluation, and especially ones returning HTTP 500 (the public PoC's invoke step returns 500 after its code has already run).
  • experiments/create requests whose artifact_location points into site-packages or a file:// path outside your normal artifact root.
  • model-versions/create requests carrying the mlflow.prompt.is_prompt tag with a relative source such as ../, followed by model-versions/get-artifact calls.
  • Experiment and registered-model names you don't recognize.

FAQ

Is my MLflow server vulnerable to CVE-2026-116692?

It is if it runs MLflow 3.12.0rc0 up to 3.17.0, a caller can reach it over the network, and either authentication is off or the caller has an account that can create experiments and upload artifacts. The exact chain also needs the server process to be able to write to its scorer package directory, which was true in the official v3.14.0 image the advisory tested.

Is CVE-2026-116692 being exploited?

Previdian, a company that runs internet sensors, reports 403 exploitation attempts against CVE-2026-116692 from 26 IP addresses between October 28 and November 11, 2026. MLflow and CISA haven't confirmed exploitation, no successful compromise has been reported, and it isn't on CISA KEV. Previdian also reports attempts on the related file-read flaw CVE-2026-116690. A harmless public proof of concept is in the GitHub advisory.

Does turning on MLflow authentication fix it?

No. It removes unauthenticated access, but any user who can create an experiment and upload artifacts can still use CVE-2026-116692, and any user who can create a registered model can still use CVE-2026-116690. Upgrading to 3.17.0 is the fix.

Is this the MLflow flaw on CISA's exploited list?

No. That's CVE-2026-64849, a server-side request forgery flaw added to KEV on August 19, 2026. This advisory covers CVE-2026-116692 and CVE-2026-116690, both fixed in 3.17.0.

What's the difference between CVE-2026-116692 and CVE-2026-116690?

CVE-2026-116692 (9.8) lets a caller run code on the server through the scorer invoke endpoint. CVE-2026-116690 (7.5) lets a caller read files the server can read through the model version get-artifact endpoint. They affect slightly different ranges (3.12.0rc0 and 3.10.0 onward) and both are fixed in 3.17.0.

Vendor
MLflow (open source)
Product
MLflow tracking server (pip package and ghcr.io/mlflow/mlflow images)
Affected versions
CVE-2026-116692: 3.12.0rc0 up to but not including 3.17.0. CVE-2026-116690: 3.10.0 up to but not including 3.17.0
Vulnerable component
/ajax-api/3.0/mlflow/scorer/invoke (RCE); model-versions/get-artifact (file read)
CVSS
9.8 (GitHub); CVE-2026-116690 is 7.5
CVSS vector
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H (GitHub); NVD has not scored either CVE yet
CWE
CWE-94, CWE-470 (code injection via scorer deserialization); CWE-22 for CVE-2026-116690
Exploitation
Reported, not on KEV
CISA KEV added
Patch
MLflow 3.17.0 (Oct 7, 2026) fixes both. SQL-backed servers need a coordinated mlflow db upgrade with writers stopped and a backup taken first

Updates

Nov 11, 2026. Status moved to Reported. Previdian reports 403 exploitation attempts against CVE-2026-116692 from 26 IPs, first seen October 28 and last seen November 11. MLflow and CISA haven't confirmed exploitation, and no compromise has been reported. Not on CISA KEV (catalog 2026.11.10).

Oct 28, 2026. First published. No exploitation of CVE-2026-116692 reported; not on CISA KEV.

Sources

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

We publish an advisory when a vulnerability is Critical or on CISA KEV, has one identifiable component to check, and has a public proof of concept or reported exploitation. Advisories are updated as exploitation status changes.