CVE-2026-116692

MLflow: Remote code execution on unauthenticated MLflow servers via third-party scorer deserialization importing caller-placed artifact modules

Published Oct 27, 2026 CVSS 3.1 · 9.8 Critical EPSS < 1% KEV · Not listed
Description
MLflow is an open source AI engineering platform for agents, large language models, and machine learning models. From 3.12.0rc0 until 3.17.0, third-party scorer deserialization in mlflow/genai/scorers/base.py accepts request-selected descendants of THIRD_PARTY_SCORER_ALLOWED_MODULES rather than requiring an exact allowed module. A caller can use MLflow artifact endpoints to place a Python module beneath a writable allowed package and submit its module and class through /ajax-api/3.0/mlflow/scorer/invoke. Scorer.model_validate_json() imports the module and constructs the class before trace processing. Authentication-disabled deployments permit this code execution without credentials, while basic-auth deployments require a principal capable of creating an experiment and run and uploading artifacts. In either configuration, the allowed scorer package must be writable by the server process. This issue is fixed in version 3.17.0.

Analysis

AI-generated

Technical impact

A vulnerability exists in the third-party scorer deserialization logic of MLflow between versions 3.12.0rc0 and 3.17.0. The deserialization process accepts any subclass of the predefined list of allowed modules without validating the specific module name. An attacker can upload a malicious Python module to an artifact path that is writable by the MLflow server and then invoke the scorer via the/​ajax‑api/​3.0/​mlflow/​scorer/​invoke endpoint. The deserialization layer imports the attacker‑supplied module and constructs the class before processing a trace, leading to arbitrary code execution with the permissions of the MLflow server process. The flaw is a classic example of insecure deserialization (CWE‑470) and code injection (CWE‑94).

Affected Systems

The affected product is MLflow, specifically versions 3.12.0rc0 through 3.16.x inclusive. Versions prior to 3.12.0rc0 and 3.17.0 and later are not impacted.

Risk and Exploitability

The CVSS score of 9.8 indicates a critical severity. The lack of an EPSS score means no current data on exploitation probability, but the vulnerability is fully exploitable on authentication‑disabled deployments and requires only minimal privilege on authentication-enabled deployments. Because the flaw allows execution of arbitrary code on the server side, the potential impact on confidentiality, integrity, and availability is total. The vulnerability is not listed in the CISA KEV catalog.

Remediation

No solution or workaround provided in the CVE record.

OpenCVE Recommended Actions

  • Apply the MLflow 3.17.0 patch or later to remove the insecure deserialization logic.
  • If a patch cannot be applied immediately, restrict artifact upload permissions so that only trusted users can write to the allowed scorer package directories.
  • Disable the scorer invocation endpoint on any unauthenticated deployments or enforce strict authentication with only users who require experiment and run creation rights.

Generated by OpenCVE AI on October 27, 2026 at 21:34 UTC.

Tracking

Sign in to view the affected projects.

Advisories

Source ID Title
Github GHSA Github GHSA GHSA-26p8-2jq9-3vq9 MLflow vulnerable to remote code execution on unauthenticated servers via third-party scorer deserialization importing caller-placed artifact modules
History
Type Values Removed Values Added
Metrics ssvc

{'options': {'Automatable': 'yes', 'Exploitation': 'poc', 'Technical Impact': 'total'}, 'version': '2.0.3'}

Type Values Removed Values Added
First Time appeared Mlflow
Mlflow mlflow
Vendors & Products Mlflow
Mlflow mlflow
Type Values Removed Values Added
Description MLflow is an open source AI engineering platform for agents, large language models, and machine learning models. From 3.12.0rc0 until 3.17.0, third-party scorer deserialization in mlflow/genai/scorers/base.py accepts request-selected descendants of THIRD_PARTY_SCORER_ALLOWED_MODULES rather than requiring an exact allowed module. A caller can use MLflow artifact endpoints to place a Python module beneath a writable allowed package and submit its module and class through /ajax-api/3.0/mlflow/scorer/invoke. Scorer.model_validate_json() imports the module and constructs the class before trace processing. Authentication-disabled deployments permit this code execution without credentials, while basic-auth deployments require a principal capable of creating an experiment and run and uploading artifacts. In either configuration, the allowed scorer package must be writable by the server process. This issue is fixed in version 3.17.0.
Title MLflow: Remote code execution on unauthenticated MLflow servers via third-party scorer deserialization importing caller-placed artifact modules
Weaknesses CWE-470
CWE-94
References
Metrics cvssV3_1

{'score': 9.8, 'vector': 'CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H'}

Weaknesses
  • CWE-470

    Use of Externally-Controlled Input to Select Classes or Code ('Unsafe Reflection')

  • CWE-94

    Improper Control of Generation of Code ('Code Injection')

Sources

Status
PUBLISHED
Assigner
GitHub_M
Published
Updated
Nov 2, 2026, 7:42 PM
Reserved
Oct 26, 2026, 8:00 PM
Status
Awaiting Analysis
Published
Oct 27, 2026, 9:18 PM
Modified
Nov 2, 2026, 8:17 PM