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

CVE-2026-2393: MLflow Has a Server-Side Request Forgery (SSRF) Vulnerability

highCVSS 3.0 7.1GHSA-65h7-c7c4-mghx

MLflow before 3.9.0 accepts any URL when you create a webhook, and the server later sends HTTP POST requests to it. An authenticated user can make the MLflow backend call internal services or cloud metadata endpoints (SSRF). Upgrade to mlflow 3.9.0 or later.

Published May 11, 2026 · Updated May 18, 2026 · Analysis Oct 27, 2026

What an attacker can do

The _create_webhook() handler in mlflow/server/handlers.py stores the url parameter without checking it. When a webhook fires, _send_webhook_request() in mlflow/webhooks/delivery.py sends a POST to that URL from the MLflow server's own network position.

An attacker needs a login that can create webhooks (the CVSS vector says low privileges, network access). They point the webhook at an internal address, such as a cloud metadata endpoint, an internal admin API or a service bound to localhost, and trigger delivery. The advisory describes cloud credential theft, internal network access and data exfiltration. The request is a POST, and the webhook test feature is another way to trigger it. Deployments on cloud VMs with instance roles carry the most risk, because the metadata service may hand out role credentials.

Who is affected

Affected: mlflow versions below 3.9.0, on any MLflow tracking server where users can create webhooks. Check the installed version:

pip show mlflow

mlflow

Ecosystem
PyPI
Affected
< 3.9.0
Fixed in
3.9.0
Check
pip show mlflow

How to fix it

Upgrade to mlflow 3.9.0 or later:

pip install --upgrade "mlflow>=3.9.0"

The fix commit makes _validate_webhook_url resolve the host with socket.getaddrinfo and reject any address that is not globally routable, including when only one of several resolved addresses is private. _send_webhook_request runs the check at delivery time. A new setting, MLFLOW_WEBHOOK_ALLOW_PRIVATE_IPS, defaults to false. The docs say to enable it only for local development. Leave it off in production.

How to tell if you were hit

List existing webhooks and check their URLs for private ranges, localhost, link-local addresses such as the metadata IP, or unfamiliar hosts. Check the MLflow server's outbound traffic and access logs on internal services for POSTs from the MLflow host that you did not expect. Review who created webhooks and when.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

A Server-Side Request Forgery (SSRF) vulnerability exists in MLflow versions prior to 3.9.0. The _create_webhook() function in mlflow/server/handlers.py accepts a user-controlled url parameter without validation, and the _send_webhook_request() function in mlflow/webhooks/delivery.py sends HTTP POST requests to this attacker-controlled URL. This allows an authenticated attacker to force the MLflow backend to send HTTP requests to internal services, cloud metadata endpoints, or arbitrary external servers. The lack of input sanitization, URL scheme filtering, or allowlist validation on the webhook URL enables exploitation, potentially leading to cloud credential theft, internal network access, and data exfiltration.

Questions

Does exploiting this need a login?

Yes. The CVSS vector requires low privileges. The attacker must be an authenticated user who can create a webhook.

What does the fix change?

In 3.9.0 the webhook URL host is resolved and rejected if any address is not globally routable. The check runs again when the request is sent. The setting MLFLOW_WEBHOOK_ALLOW_PRIVATE_IPS can turn the check off.

Can it read responses from internal services?

The advisory lists data exfiltration and credential theft as possible outcomes. It does not detail what response data returns to the attacker, so treat internal services as reachable.

Is it exploited in the wild?

No source says so, and the entry is not in CISA's KEV catalog.

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