MLflow unauthenticated SSRF: affected servers and fixes
MLflow has reported a critical vulnerability related to SSRF request forgery affecting versions prior to 3.15.0 of its platform. This flaw would allow an attacker to interact with the model registry's webhook API without authentication, as reported by INCIBE-CERT on August 18, 2026.
Technical details of the vulnerability
The vulnerability, identified as CVE-2026-64849, is found in the unauthenticated POST endpoint /api/2.0/mlflow/webhooks/{id}/test. In versions prior to 3.15.0, the function _validate_webhook_url() only validated the original URL, but the webhook delivery process followed redirects without restricting hostname resolution. This allowed attackers to perform SSRF requests and access internal or cloud services, obtaining information about the status and content of responses.
Implications of SSRF request forgery in MLflow
The issue lies in the fact that the validation mechanism did not block HTTP redirects or DNS rebinding attacks, making exploitation easier. By not requiring authentication, SSRF request forgery could be used to access internal resources through the vulnerable server. This poses a significant security risk for infrastructures using earlier versions of MLflow.
Recommended measures and update
To mitigate this vulnerability, MLflow has released version 3.15.0, which fixes the flaw in the validation and webhook delivery process. The update ensures that redirects and DNS resolutions are properly controlled to prevent unauthorized access. MLflow users are advised to upgrade to this version to avoid potential SSRF request forgery attacks.
Context and scope of the alert
MLflow is an open-source platform focused on artificial intelligence engineering and machine learning models. According to INCIBE, the severity of this flaw is critical due to the possibility of unauthenticated access to internal resources through SSRF techniques.
This alert highlights the need to keep development and operational tools updated in AI environments to avoid risks arising from vulnerabilities in integration and automation components.
More information and technical details can be found in the original source provided by INCIBE.
How to check whether your MLflow server is exposed
The risk concerns an MLflow Tracking Server whose webhook API can be reached. Using MLflow solely as a client library does not itself mean that an exposed vulnerable server exists. Check whether the webhook test endpoint is reachable and whether the running server uses an affected release. CVE-2026-64849 allows a webhook test request to reach internal destinations after a redirect or a fresh DNS lookup.
This advisory does not establish that every MLflow installation has been attacked or that the issue grants arbitrary system execution. The documented impact is the ability to read responses from services reachable by the affected host, potentially including internal services. CISA's inclusion of the CVE in its Known Exploited Vulnerabilities catalogue makes prompt investigation important.
A practical administrator checklist
- Inventory: identify the version actually running, not merely the version listed in a dependency file. Include containers, replicas and accessible test environments.
- Patch: upgrade to MLflow 3.15.0 or later following the vendor's guidance. Confirm the deployed version after a restart or redeployment.
- Reduce exposure: restrict access to the Tracking Server and webhook API using suitable network and authentication controls while remediation is under way. URL validation alone should not be assumed to stop redirects.
- Investigate: if the server was reachable, review available webhook access logs and unusual requests towards internal destinations. No automated alert is not proof that no abuse occurred.
Consult the MLflow security advisory and the Canadian Centre for Cyber Security notice for scope and release details. Editorial review: 11 November 2026. This article does not certify the security or compromise status of any particular installation.
Updated: 11/11/2026.
This article may have been prepared with the assistance of artificial intelligence tools. View AI policy
You can support the project or share this article in one click. At least this block does something useful.