NewApp Monitoring is liveSign up free
AgentbeamBeta
AI agent vulnerability · MLflow

CVE-2023-6909: MLflow Path Traversal Vulnerability

highCVSS 4.0 8.7 · 3.1 7.5GHSA-5r3q-93q3-f978

MLflow before 2.9.2 can be tricked into reading files outside the artifact root with an encoded path. An unauthenticated network caller can read files the MLflow server process can read. Fixed in mlflow 2.9.2. The EPSS score is very high, near the 99.8th percentile.

Published Dec 20, 2023 · Updated Sep 24, 2024 · Analysis Nov 2, 2026

What an attacker can do

MLflow checks artifact paths in validate_path_is_safe in mlflow/utils/uri.py. The check ran on the raw string. A path with URL-encoded traversal such as %2E%2E%2F passed the check and was decoded later, so it escaped the artifact directory. The fix adds urllib.parse.unquote before validation and makes the error read just "Invalid path", without echoing the input.

The attacker needs only network access to the MLflow tracking server, for example the mlflow-artifacts endpoints. No login is needed, because MLflow has no authentication by default. The CVSS vector scores confidentiality high and nothing else, so the outcome is file disclosure. On an ML host that means model files, training data, config files, .env files and cloud credentials, anything the server user can read.

A sibling flaw fixed in the same release, CVE-2023-6976, covers writes through the same family of path checks. A public weakness in a server that has no built-in login is a likely target, and the EPSS score reflects that. I found no source stating confirmed attacks.

Who is affected

mlflow on PyPI below 2.9.2, on any tracking server reachable by untrusted users.

Check the installed version:

pip show mlflow

For a containerized server, run the same command inside the container.

mlflow

Ecosystem
PyPI
Affected
< 2.9.2
Fixed in
2.9.2
Check
pip show mlflow

How to fix it

Upgrade to mlflow 2.9.2:

pip install --upgrade "mlflow>=2.9.2"

Restart the server. The fix commit decodes the path before validating it. The sources give no workaround. Until you upgrade, put the server behind authentication on a reverse proxy and keep it off untrusted networks.

How to tell if you were hit

Search access logs of the tracking server for requests to artifact endpoints that contain %2E%2E, %2e%2e or %2F sequences, or ../ in the path. On patched versions such requests return a 400 with the message Invalid path. Requests like these to an unpatched server that returned 200 mean files may have been read. Rotate secrets stored on the host if you find them.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

Path Traversal: '\..\filename' in GitHub repository mlflow/mlflow prior to 2.9.2.

Questions

Is mlflow 2.9.2 affected by CVE-2023-6909?

No. 2.9.2 is the patched version. Versions below it are vulnerable.

Can the attacker write or delete files with this?

Not with this flaw. The CVSS vector scores only confidentiality as high. A related issue, CVE-2023-6976, covers writing files.

Why is the EPSS score so high?

The data shows a score of about 0.90, near the 99.8th percentile. The flaw needs no login and no user interaction, and MLflow servers have no login by default. The sources do not say it is exploited in the wild.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology

More MLflow vulnerabilities

All MLflow entries