What is MLflow CVE-2026-64849?
MLflow's security advisory describes an unauthenticated server-side request forgery in webhook delivery. On a default MLflow server, an attacker can call POST /api/2.0/mlflow/webhooks/{id}/test to reach internal or cloud-metadata services such as 169.254.169.254, and the response status and body come back to the attacker. The URL check could be bypassed with HTTP redirects and DNS rebinding.
The fix shipped in MLflow 3.15.0, and the advisory says IPv6-transition bypasses of that fix were closed in 3.16.0. CISA added CVE-2026-64849 to its Known Exploited Vulnerabilities catalog on August 19, 2026, with a September 2 due date. GitHub rates it CVSS 3.1 9.3.
Which MLflow versions are affected?
- MLflow 3.3.0 up to, but not including, 3.15.0.
- IPv6 and NAT64 deployments should run 3.16.0 or later.
- Default
mlflow serverdeployments, where webhook endpoints are unauthenticated unless the optional auth plugin is loaded.
How to check if you're affected
- Run
pip show mlflowormlflow --versionon every tracking server. - Check whether the tracking server is reachable from untrusted networks and whether the auth plugin is loaded.
- Review logs for webhooks created through
POST /api/2.0/mlflow/webhooksand calls to the webhook test endpoint.
What to do now
- Upgrade to MLflow 3.15.0 or later, or 3.16.0 or later on IPv6 and NAT64 networks.
- Leave
MLFLOW_WEBHOOK_ALLOW_PRIVATE_IPSset to false outside local development. - Put the tracking server behind authentication, and block its access to cloud metadata endpoints where you can.
Why SSRF in ML tooling is a cloud credential risk
ML platforms usually run inside cloud accounts with roles that can read data and models. An SSRF that reaches the instance metadata service can hand those role credentials to an attacker. Require authentication on internal tools, use metadata service protections such as IMDSv2 hop limits, and validate webhook targets after every redirect, not just the first URL.
Where Nullify fits
How Nullify Covers This Class of Attack
Nullify moves actively exploited vulnerabilities in the dependencies and images you use up the backlog as soon as they're disclosed, and prepares the version change as a pull request for each owner to review.
The vulnerable version is set in your dependency manifests, lockfiles or images, which is where to look.
Frequently Asked Questions
Which MLflow versions fix CVE-2026-64849?
MLflow 3.15.0 fixes it. The advisory says bypasses on IPv6 and NAT64 networks were fixed in 3.16.0, so use 3.16.0 or later there. Versions 3.3.0 up to 3.15.0 are affected.
How is MLflow CVE-2026-64849 exploited?
An unauthenticated attacker uses MLflow's webhook test endpoint to send requests that redirect to internal or cloud-metadata addresses, and MLflow returns the response. That can expose cloud credentials from the metadata service.
Is CVE-2026-64849 exploited in the wild?
Yes. CISA added CVE-2026-64849 to its Known Exploited Vulnerabilities catalog on August 19, 2026, which means it has evidence of active exploitation. It gave US federal civilian agencies until September 2 to fix it. Treat exposed MLflow instances as urgent.
Update Log
- Nov 17, 2026: page published from the primary sources linked above.