⚠ Reference material — not professional advice. Test in staging, back up first, verify against your specific version. Use your own judgment for your environment.
● High · CVSS 7

How to Fix CVE-2026-4137: Deserialization RCE in mlflow/mlflow

Other vulnerabilities in the same area that are worth patching alongside this one:

*By Sai Kiran Pandrala*

⚡ At a glance
SeverityCVSS 7 - High
Actively exploited?Not currently listed in CISA KEV
Affectedunspecified < 3.11.0
Fixed inSee vendor advisory
Type (CWE)CWE-378: Creation of Temporary File With Insecure Permissions

What is CVE-2026-4137?

CVE-2026-4137 is an unsafe deserialization in mlflow/mlflow. The application accepts attacker-controlled serialized objects and reconstructs them without validating their type, so a crafted payload triggers code execution inside the running process. Unauthenticated remote code execution is the typical impact. Vendor description: 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().

Why this CVE matters

Deserialization bugs are a favorite of ransomware operators because they convert a single HTTP request into full code execution on the target host. Public proof-of-concept code for this CVE class typically appears within days of disclosure, and weaponized exploits follow shortly after.

For deployments of mlflow/mlflow that have been exposed to the public internet during the disclosure window, the operating assumption should be that scanning has already happened. Even where exploitation has not been publicly observed, scanning for the vulnerable fingerprint is cheap and routine. Patching closes the door; log review and credential rotation close out the rest of the response.

Am I affected?

You are affected if your installation matches any of these version ranges:

Check your installed version against the list above. If you cannot determine the version, treat the system as affected and follow the upgrade path below.

Open mlflow/mlflow's About dialog or run the vendor-documented version-check command. Compare the result against the affected ranges in the advisory.

How to fix CVE-2026-4137

  1. Read the vendor advisory in full: https://huntr.com/bounties/648dc30b-76c7-4433-86b8-f43d926fd8d6
  2. Upgrade mlflow/mlflow to the patched build listed in the vendor advisory.
  3. Back up the configuration (and database, where applicable) before upgrading.
  4. Rotate any credentials, API keys, or session tokens that the vulnerable service touched. An unauthenticated RCE-class flaw means anything the process could see should be treated as exposed.
  5. Apply the patch in a maintenance window. For HA pairs, upgrade the standby node first, fail over, then upgrade the former primary.
  6. Restart the affected service so the patched binary loads, then verify the new version (see verification section).

Update the Python package


# CVE-2026-4137 affects mlflow/mlflow unspecified < 3.11.0. Fixed in 3.11.0.
# Vendor advisory: https://huntr.com/bounties/648dc30b-76c7-4433-86b8-f43d926fd8d6

# 1. Show the currently installed version.
python -m pip show mlflow-mlflow | grep -i version

# 2. Upgrade to the patched release.
python -m pip install --upgrade "mlflow-mlflow>=3.11.0"

# 3. For projects pinned via requirements.txt, bump the pin and re-sync.
sed -i 's/^mlflow-mlflow==.*/mlflow-mlflow==3.11.0/' requirements.txt
python -m pip install -r requirements.txt

# 4. Verify.
python -m pip show mlflow-mlflow | grep -i version


# Same flow on Windows.
python -m pip install --upgrade "mlflow-mlflow>=3.11.0"
python -m pip show mlflow-mlflow

Verify the fix landed


# CVE-2026-4137 verification checklist.

# 1. Confirm the running version matches 3.11.0 (replace the version probe with
#    the platform-specific command shown above).

# 2. Re-scan the host with your vulnerability scanner (Nessus, Qualys, Tenable,
#    OpenVAS, Wazuh). The scanner must no longer flag CVE-2026-4137.

# 3. Inspect recent service and kernel logs for crash-loops or rollback events.
journalctl -u <service-name> --since "10 minutes ago"
dmesg --since "10 minutes ago"

# 4. Cross-check the running build against the vendor advisory:
#    https://huntr.com/bounties/648dc30b-76c7-4433-86b8-f43d926fd8d6

If you cannot patch immediately

There is no safe runtime mitigation for deserialization flaws beyond removing exposure: block the affected endpoint at a reverse proxy or WAF and restrict access to authenticated, trusted users only. Patch as soon as possible.

How to verify the fix worked

If your installation was internet-reachable during the disclosure window, treat log review as part of the remediation rather than an optional follow-up. Look for unexpected administrator accounts in mlflow/mlflow, scheduled tasks or cron jobs you did not create, new files in web-accessible directories, and outbound connections to addresses not in your baseline. Suspicious requests to the vulnerable endpoint immediately followed by successful 200-class responses with unusually large bodies are a strong indicator of exploitation.

Frequently asked questions

Is CVE-2026-4137 being exploited in the wild?

Public exploitation has not been confirmed by CISA at the time of writing. Treat the patch as time-sensitive anyway; reports often lag actual abuse.

Will a WAF or IDS rule fully mitigate CVE-2026-4137?

No. Network-layer filters can reduce noise and slow opportunistic scanners, but they will not stop a determined attacker. The vendor patch is the only durable fix.

Do I need to assume compromise if my mlflow/mlflow was internet-facing and unpatched?

For an unauthenticated RCE-class flaw exposed to the public internet during the known exploitation window, yes. Review logs, rotate credentials the process could access, and look for unexpected accounts, scheduled tasks, or outbound connections.

References

*This guide was assembled from the official vendor advisory, the NVD record, and the CISA KEV catalog entry on 2026-05-25. Always confirm against the vendor advisory before applying changes in production.*