Description

A critical, unauthenticated server-side request forgery (SSRF) vulnerability in MLflow, identified as CVE-2026-64849, is reportedly being exploited in the wild shortly after its disclosure, according to watchTowr. The issue impacts MLflow releases earlier than 3.15.0 and could allow remote attackers to access cloud metadata, internal services, credentials, and other sensitive information. The vulnerability is associated with MLflow’s model-registry webhook testing feature. On a default Tracking Server, the POST /api/2.0/mlflow/webhooks/{id}/test endpoint can be accessed without authentication, enabling an attacker to initiate webhook requests and potentially obtain the resulting response status and content. The underlying weakness stems from insufficient validation when following HTTP redirects. Although MLflow introduced protections in version 3.10.0 to block webhook requests to private and reserved IP addresses, the validation is performed against the destination supplied initially. If that publicly accessible destination responds with a redirect, MLflow can follow it without applying the same validation to the redirected address. An attacker could abuse this behavior by directing requests toward sensitive endpoints, including cloud metadata services such as 169.254.169.254 or internal resources bound to loopback addresses such as 127.0.0.1. Successful exploitation could provide access to temporary cloud credentials, instance metadata, internal application responses, configuration information, or secrets available to the vulnerable server. watchTowr reports that its global honeypot infrastructure detected exploitation attempts targeting cloud-hosted MLflow deployments soon after the vulnerability was disclosed, with observed activity including attempts to obtain credentials and secrets. Organizations should treat internet-accessible MLflow instances running versions older than 3.15.0 as potentially exposed and prioritize upgrading to MLflow 3.15.0 or later. Administrators should also restrict direct internet access to Tracking Servers, enforce authentication through a reverse proxy or identity-aware gateway, review webhook configurations, and examine logs for suspicious webhook-test requests, redirect activity, metadata-service access, or unusual outbound connections. Where compromise is suspected, potentially exposed cloud credentials and secrets should be rotated and the affected host investigated for unauthorized access.