CVE-2026-116692: Unauthenticated RCE in MLflow via Third-Party Scorer Deserialization
What's Happening
CVE-2026-116692 is a critical (CVSS 9.8) remote code execution vulnerability in MLflow's tracking server. The root cause is twofold: unauthenticated deserialization of third-party "scorers" (CWE-94, code injection) combined with unsafe reflection that imports caller-supplied artifact modules (CWE-470). An attacker who can reach an exposed MLflow tracking server can upload a malicious Python module as a run artifact, then trigger the server to import and execute it — no authentication required.
Affected Software
The vulnerability affects mlflow (distributed via pip) versions >= 3.12.0rc0 and < 3.17.0. Any internet- or network-exposed MLflow tracking server running an affected version without authentication in front of it is vulnerable.
Exploitation Status
A public proof-of-concept exists for this vulnerability. There is no indication yet that it has been added to CISA's KEV catalog or is being mass-exploited in the wild, but public PoC code significantly lowers the bar for opportunistic attackers. Defenders should treat internet-exposed, unpatched MLflow instances as an active risk and prioritize patching or network isolation rather than waiting for confirmed in-the-wild exploitation.
How Our Detection Catches It
This detection ships across Microsoft Sentinel (KQL), Splunk (SPL), Elastic (EQL), QRadar (AQL), Sumo Logic, Chronicle (YARA-L), and CrowdStrike (CQL). The logic correlates three signals that are each individually unusual but highly indicative in combination:
- Anomalous artifact uploads to the MLflow REST API, particularly uploads of Python module files to run/experiment artifact paths
- Requests to scorer or evaluation endpoints originating from unauthenticated or unexpected clients
- Child-process spawning from the MLflow server process (or its underlying gunicorn worker) immediately following an artifact upload or evaluation call — the telltale sign of server-side code execution
Flagging this sequence rather than any single event lets the rule catch exploitation attempts with low false-positive volume, since legitimate MLflow usage rarely produces process spawns tied directly to artifact ingestion.
Full Detection Details
For the complete KQL/SPL/EQL/AQL queries, investigation guidance, and atomic test procedures, see the full detection page: CVE-2026-116692 on df00tech.