Summary
MLflow versions <=3.10.1.dev0 expose multipart upload (MPU) artifact endpoints without enforcing resource-level authorization checks when --serve-artifacts is enabled.
An attacker who can reach these endpoints may overwrite artifacts owned by other users, enabling cross-tenant model tampering and potential remote code execution when poisoned models are later loaded.
Affected context
This vulnerability appears in MLflow deployments that:
- Run MLflow with the
--serve-artifactsoption enabled - Expose
/mlflow-artifacts/mpu/*endpoints over a network (intranet or internet) - Rely on MLflow's built-in authorization to protect multi-user or multi-tenant artifact stores
It is particularly relevant in shared ML platforms, MLOps pipelines, or model registries where multiple users or services upload and consume artifacts through the same MLflow server.
Exploit preconditions
- Attack surface: HTTP(S) endpoints under
/mlflow-artifacts/mpu/*on the MLflow server. - Required access: Network access to the MLflow artifact service; depending on deployment, this may be internal, VPN-only, or internet-exposed.
- Trigger point: Initiating multipart uploads or completing MPU operations that target artifacts associated with other users or projects.
- Why exploitation is possible: The MPU endpoints perform authorization checks at a coarse level and fail to enforce fine-grained, resource-level permissions for the specific artifact being written, allowing unauthorized cross-user writes.
Why this happens
The underlying issue is incomplete enforcement of authorization on sensitive artifact operations.
While the MLflow server may authenticate requests or apply some high-level access control, the /mlflow-artifacts/mpu/* endpoints do not correctly verify that the caller is allowed to write to the particular artifact path or resource.
This maps to CWE-862: Missing Authorization, where the software either omits or inconsistently applies permission checks, especially on specialized sub-resources such as multipart upload control paths.
Abstractly, it represents an insecure pattern where secondary or "helper" endpoints bypass the same fine-grained policy used by primary artifact APIs.
Potential impact
- Attackers may overwrite or replace model artifacts owned by other users.
- Adversaries could inject malicious model files or supporting artifacts, leading to model supply chain poisoning.
- Compromised or backdoored models can enable arbitrary code execution when loaded by downstream applications or pipelines that execute model initialization or custom code.
- Integrity of experiments, training results, and audits can be undermined through silent artifact tampering.
- Cross-tenant data isolation may be broken in shared ML platforms.
General mitigation guidance
Effective mitigation typically centers on enforcing consistent, least-privilege authorization for all artifact operations and reducing the trust placed in network-accessible artifact endpoints.
Useful defensive patterns include:
- Applying uniform, resource-level authorization checks to every artifact-related endpoint, including multipart upload helpers.
- Designing authorization middleware that is applied centrally to all routes, avoiding one-off or ad hoc access checks in individual handlers.
- Segmenting artifact storage per tenant or project with separate credentials or namespaces, so that a compromise in one context cannot trivially impact others.
- Restricting network exposure of
--serve-artifactsto trusted internal networks or dedicated service accounts rather than broad user access. - Validating and scanning models and artifacts before they are promoted to production, treating them as untrusted inputs even when retrieved from internal MLflow servers.
How teams detect issues like this
Organizations can look for this class of issue and signs of exploitation using:
- Static and configuration analysis focused on route handlers and decorators, verifying that all artifact endpoints (including
/mlflow-artifacts/mpu/*) are wrapped in the same authorization logic as primary APIs. - Code review checklists that explicitly call out multipart upload or "utility" endpoints as high-risk for missing access control.
- Logging and monitoring of artifact write operations, especially cross-user or cross-project writes, with anomaly detection on unexpected owners or paths.
- Policy enforcement in CI/CD that prevents deployment of services where security middleware is not uniformly attached to all HTTP routes.
- Periodic access-control testing of MLflow environments, including attempts to write to artifacts belonging to other users from constrained accounts.

