Summary
MLflow can allow an authenticated user to create a model version that references another user's artifact directory without the required read permission. When the referenced artifact is retrieved, this authorization gap may expose files from a different user's model or run.
Affected context
The issue occurs in MLflow deployments where model-version creation accepts a run or model reference, and later artifact retrieval relies on path containment without fully verifying authorization for the referenced resource. The risk is relevant when authenticated users have access to model-version APIs but should not be able to read artifacts owned by other users.
Exploit preconditions
- Attack surface: MLflow model-version creation and artifact retrieval APIs
- Required access: An authenticated MLflow account with permission to create model versions
- Trigger point: A model version references an artifact directory associated with another user's run or model
- Why exploitation is possible: Path containment checks confirm that a path is within an artifact location but do not establish that the requesting user has READ permission for the referenced resource
Why this happens
The underlying bug is an authorization decision that validates filesystem or artifact-path relationships without enforcing resource ownership or access permissions. This maps to CWE-862: Missing Authorization, because the application permits an authenticated principal to access a protected artifact without confirming the required READ authorization for that specific resource.
Potential impact
- Unauthorized users may read model artifacts or related files
- Sensitive datasets, model binaries, configuration, or metadata could be disclosed
- Confidentiality of projects belonging to other users may be reduced
- Exposed artifacts could support further attacks against connected ML infrastructure
General mitigation guidance
Mitigation generally focuses on enforcing authorization at every resource access boundary rather than relying only on path validation.
Common defensive patterns include:
- Verify READ permission for the referenced run, model, and artifact before creating or serving a model version
- Bind authorization decisions to the authenticated principal and the canonical resource identity
- Treat user-supplied run and model references as untrusted identifiers
- Apply deny-by-default access control for cross-user and cross-tenant artifact access
- Separate artifact storage permissions from model-version metadata permissions
- Run services with least-privilege access to artifact stores and avoid exposing shared storage paths directly
How teams detect issues like this
Organizations can identify similar authorization gaps through:
- Static analysis and code review of handlers that validate paths without checking ownership or permissions
- Tests that attempt cross-user access to runs, models, model versions, and artifacts
- Policy checks requiring authorization enforcement at both object-creation and object-retrieval boundaries
- Audit-log monitoring for model-version references and artifact reads crossing user or tenant boundaries
- Automated security review tools that flag resource access based solely on identifier or path validation

