CVE-2026-4137: MLFlow Creates a Temporary File With Insecure Permissions
MLflow before 3.11.0 creates temporary model directories with loose permissions: 0o777 in get_or_create_nfs_tmp_dir() and 0o770 in _create_model_downloading_tmp_dir() (CVE-2026-4137). A local user can tamper with cloudpickle model artifacts and get code execution when they load. Upgrade to mlflow 3.11.0.
Published May 18, 2026 · Updated Sep 1, 2026 · Analysis Nov 3, 2026
What an attacker can do
get_or_create_nfs_tmp_dir() in mlflow/utils/file_utils.py makes a world-writable directory (0o777). _create_model_downloading_tmp_dir() in mlflow/pyfunc/__init__.py makes a group-writable one (0o770). Model artifacts are staged in these directories before loading. Some are cloudpickle-serialized Python objects, and cloudpickle.load() runs code embedded in the file (CWE-378).
A local attacker who can write to the directory swaps a staged artifact for a malicious one. When the victim's MLflow process loads the model, the attacker's code runs with that process's rights. The CVSS 3.0 vector is local, high complexity, low privileges, no user interaction.
The advisory calls this especially serious in shared NFS setups such as Databricks, where NFS is enabled by default and many users can reach the same temp mount. On a single-user machine the risk is small. On shared training or serving hosts, a low-privileged user could run code as the account that serves or trains models and read its credentials and data. The issue continues the class fixed in CVE-2025-10279, which the description says was only partly fixed.
Who is affected
mlflow from 1.26.0 up to but not including 3.11.0. Shared hosts, multi-user notebook clusters and any setup with an NFS temp directory are the real exposure. Check the version:
pip show mlflowAlso look at the modes of the MLflow temp directories on shared systems with ls -ld.
mlflow
- Ecosystem
- PyPI
- Affected
- >= 1.26.0, < 3.11.0
- Fixed in
- 3.11.0
- Check
pip show mlflow
How to fix it
Upgrade to mlflow 3.11.0:
pip install --upgrade "mlflow>=3.11.0"The fix commit changes the NFS temp directory mode from 0o777 to 0o750 in get_or_create_nfs_tmp_dir() and the model download directory from 0o770 to 0o750 in _create_model_downloading_tmp_dir(). The sources list no other workaround. Until you upgrade, avoid loading models on hosts where untrusted local users share the temp or NFS location.
How to tell if you were hit
On shared hosts, list the MLflow temporary and NFS directories and flag any with mode 777 or 770 and unexpected group members. Look for artifact files, especially pickle or cloudpickle files, whose owner or modification time does not match the account that created the model. Check for unexpected child processes spawned by the MLflow serving or training process after a model load.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
In mlflow/mlflow versions prior to 3.11.0, the get_or_create_nfs_tmp_dir() function in mlflow/utils/file_utils.py creates temporary directories with world-writable permissions (0o777), and the _create_model_downloading_tmp_dir() function in mlflow/pyfunc/__init__.py creates directories with group-writable permissions (0o770). These insecure permissions allow local attackers to tamper with model artifacts, such as cloudpickle-serialized Python objects, and achieve arbitrary code execution when the tampered artifacts are deserialized via cloudpickle.load(). This vulnerability is particularly critical in environments with shared NFS mounts, such as Databricks, where NFS is enabled by default. The issue is a continuation of the vulnerability class addressed in CVE-2025-10279, which was only partially fixed.
Credited to Chi-Song-Owen.
Questions
- Is mlflow 3.10 affected by CVE-2026-4137?
Yes. The vulnerable range is 1.26.0 up to but not including 3.11.0.
- Can this be exploited remotely?
The CVSS vector is local. The attacker needs an account or process on the same host or shared NFS mount and write access to the temp directory.
- Why are Databricks users singled out?
The description says NFS is enabled by default there, and shared mounts make the world-writable directory reachable by other users.
- How does it relate to CVE-2025-10279?
The description says it continues that vulnerability class, which was only partially fixed.
References
Read for this analysis:
- https://nvd.nist.gov/vuln/detail/CVE-2026-4137
- https://huntr.com/bounties/648dc30b-76c7-4433-86b8-f43d926fd8d6
- https://github.com/advisories/GHSA-4x5p-f36r-mxxr
- https://github.com/mlflow/mlflow/commit/bd8854fb2b4b168a4384ebddc6c4a0b5d56aff01
- https://github.com/pypa/advisory-database/tree/main/vulns/mlflow/PYSEC-2026-2655.yaml
- https://github.com/advisories/GHSA-f2m9-wcf4-cwwx
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 entriesCVE-2026-116690
MLflow vulnerable to arbitrary file read on unauthenticated servers via prompt model version source
high7.5Oct 27, 2026CVE-2026-116691
MLflow CrewAI autologging records LLM provider API keys in logged trace spans
medium6.5Oct 27, 2026CVE-2026-116692
MLflow vulnerable to remote code execution on unauthenticated servers via third-party scorer deserialization importing caller-placed artifact modules
critical9.8Oct 27, 2026CVE-2026-115886
MLflow: SSRF on unauthenticated MLflow servers via third-party scorer kwargs in the scorer invoke, register, and GenAI evaluate endpoints
medium6.5Oct 23, 2026CVE-2026-53995
MLflow FileStore trace IDs allow path traversal and arbitrary file writes
high8.2Oct 20, 2026GHSA-gqvg-gmmx-x4hm
MLFLOW_ALLOW_PICKLE_DESERIALIZATION=False safety control bypassed by mlflow.statsmodels flavor — RCE via crafted model artifact
high8.8Sep 1, 2026
