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

CVE-2026-2734: MLflow authenticated users can enumerate any registered model versions due to lack of per-model permissions checks

mediumCVSS 3.0 6.5GHSA-w5xq-c4pf-ghq7

MLflow up to 3.9.0 with basic-auth enabled does not filter model version searches by permission. Any logged-in user can list every model version across all registered models through the `SearchModelVersions` REST endpoint or the `mlflowSearchModelVersions` GraphQL query. Upgrade to mlflow 3.10.0.

Published May 21, 2026 · Updated Jun 12, 2026 · Analysis Oct 26, 2026

What an attacker can do

With basic-auth on, MLflow checks requests through validator and handler tables. SearchModelVersions was missing from BEFORE_REQUEST_VALIDATORS and AFTER_REQUEST_HANDLERS for REST, and from GraphQLAuthorizationMiddleware.PROTECTED_FIELDS for GraphQL (CWE-284). Search results were therefore never filtered by the permissions on each version's parent registered model.

The attacker needs any valid account on the server, even one with no READ right on a model. They call the search endpoint, or the GraphQL query, and get all versions of all models.

The returned records carry model names, version descriptions, source URIs, tags and other metadata. On a multi-tenant server this can reveal proprietary or confidential details, such as unreleased model names, internal storage paths in source URIs and how teams label their work. The vector rates confidentiality as high and integrity and availability as none, so the flaw reads metadata and does not change anything. It can help an attacker plan further steps, for example by finding artifact locations.

Who is affected

The mlflow package on PyPI before 3.10.0, on servers that run basic-auth (--app-name basic-auth) and hold models of different users or teams. Servers with a single trusted group see little impact. Check:

pip show mlflow

Check the server's launch command for --app-name basic-auth and whether GraphQL is reachable.

mlflow

Ecosystem
PyPI
Affected
< 3.10.0
Fixed in
3.10.0
Check
pip show mlflow

How to fix it

Upgrade to mlflow 3.10.0:

pip install --upgrade 'mlflow>=3.10.0'

The fix commit adds filter_search_model_versions, a REST response filter registered for SearchModelVersions. Admins are skipped. For everyone else it removes versions whose parent registered model the user cannot read. For GraphQL, mlflowSearchModelVersions is added to the protected fields and its results go through the same read check. The search still succeeds and returns only the readable subset, so users see fewer results, not an error.

The sources give no workaround. Until you upgrade, treat model metadata as visible to every account on the server.

How to tell if you were hit

The sources give no log signatures. Check the access logs for calls to the model version search endpoint and to GraphQL from accounts that have no reason to browse the registry, especially calls with no filter that page through everything. Compare which users searched with which users hold permissions on the models. If sensitive source URIs or descriptions were exposed, review them for secrets and rotate any you find.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

In mlflow/mlflow versions up to 3.9.0, the SearchModelVersions REST API endpoint and the mlflowSearchModelVersions GraphQL query lack proper per-model authorization checks when basic authentication is enabled. This allows any authenticated user to enumerate all model versions across all registered models, regardless of their permission level. The issue arises due to the absence of SearchModelVersions in the BEFORE_REQUEST_VALIDATORS and AFTER_REQUEST_HANDLERS for the REST API, and its omission from GraphQLAuthorizationMiddleware.PROTECTED_FIELDS for GraphQL. This vulnerability can expose sensitive information such as model names, version descriptions, source URIs, tags, and other metadata, potentially revealing proprietary or confidential details in multi-tenant environments. The issue is resolved in version 3.10.0.

Questions

What can a low-privileged user see?

All model versions of all registered models: names, descriptions, source URIs, tags and other metadata, regardless of their permissions.

Does it expose model files?

The advisory covers metadata returned by search. It says nothing about downloading artifacts through this flaw.

Which version fixes CVE-2026-2734?

mlflow 3.10.0. Search results are then filtered by the user's READ permission on each parent model.

Does it affect servers without basic-auth?

The flaw is in the basic-auth permission layer. Without it there is no per-model permission to bypass.

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