CVE-2026-64849

ZDI-26-846 · ZDI-CAN-31844 · published 17 Aug 2026 · updated 20 Nov 2026

MLflow: Unauthenticated full-read SSRF in webhook delivery: _validate_webhook_url bypassed via unvalidated HTTP redirects (and DNS rebinding)

10.0CriticalCVSS 9.3EPSS 11%CISA KEV · due 2 Sep 20263 public exploits

Exploited in the wild, in CISA KEV since 19 Aug 2026 (due 2 Sep 2026), with 3 public exploits. Fixed in 3.15.0.

MLflow is an open source AI engineering platform for agents, large language models, and machine learning models. Starting in 3.3.0 and prior to 3.15.0, the POST /api/2.0/mlflow/webhooks/{id}/test endpoint (accessible without credentials when server authentication is not enabled) calls _validate_webhook_url() in mlflow/utils/validation.py only for the original URL while mlflow/webhooks/delivery.py follows redirects and re-resolves the hostname without pinning the validated address, allowing attackers to reach internal or cloud metadata services and receive response_status and response_body. The unauthenticated attack scenario described here applies to deployments without authentication. With MLflow’s built-in basic authentication enabled, requests to this endpoint require valid credentials; this issue does not itself bypass authentication. This issue is fixed in version 3.15.0.

Briefing

Written from the sources below · claude-haiku-5-5

MLflow versions from 3.3.0 before 3.15.0 contain an unauthenticated server-side request forgery in the webhook test endpoint. Webhook delivery validates only the original URL, so redirects and DNS rebinding can reach internal or cloud metadata services, returning response status and body to the attacker.

Why it mattersListed in CISA KEV as known exploited, with public nuclei template material available; no authentication is required, so exposed MLflow servers are at immediate risk.

ActionUpgrade MLflow to 3.15.0 or later; until then, restrict access to the /api/2.0/mlflow/webhooks endpoints and block outbound requests from the MLflow host to internal and metadata addresses.

Affected

Vendors, products and packages with version ranges

VendorProductVersionsFixed
MLflowMLflow>=3.3.0 <3.15.0
lfprojectsmlflow>=3.3.0 <3.15.03.15.0

Exploit evidence

3 public sources

kev

Timeline

Every event on this vulnerability

PublishedCVE-2026-64849·
MLflow versions from 3.3.0 before 3.15.0 contain an unauthenticated server-side request forgery in the webhook test endpoint. Webhook delivery validates only the original URL, so redirects and DNS rebinding can reach internal or cloud metadata services, returning response status and body to the attacker.
7.0High
Public exploitCVE-2026-64849·
MLflow versions from 3.3.0 before 3.15.0 contain an unauthenticated server-side request forgery in the webhook test endpoint. Webhook delivery validates only the original URL, so redirects and DNS rebinding can reach internal or cloud metadata services, returning response status and body to the attacker.
10.0Critical

Related

Other vulnerabilities in MLflow, strongest first

Embed

A live severity badge for an advisory, README or ticket

CVE-2026-64849 severity on Zeptric[![CVE-2026-64849](https://zeptric.io/badge/cve/CVE-2026-64849.svg)](https://zeptric.io/cve/CVE-2026-64849)