Summary
MLflow's trace API endpoints lack experiment-level authorization checks when authentication is enabled, allowing any logged-in user to access trace data across experiments.
Under these conditions, users can read, delete, or modify traces for experiments they are not authorized to access, potentially exposing sensitive information and corrupting audit trails.
Affected context
This vulnerability occurs in MLflow deployments that:
- Use versions prior to 3.14.0
- Run with authentication enabled and multiple users or services
- Rely on experiment-level authorization policies for isolation of trace data
In such setups, trace endpoints are exposed through the MLflow tracking server, but are not consistently protected by the same authorization layer that governs experiments and runs.
Exploit preconditions
- Attack surface: MLflow tracking server trace API endpoints
- Required access: Any authenticated MLflow user account or API token
- Trigger point: Invoking trace-related operations (read, update, delete) on experiments where the caller lacks experiment-level permissions
- Why exploitation is possible: The
_before_requesthandler does not register authorization validators for trace endpoints, so requests bypass experiment-level authorization checks and are processed as if fully permitted.
Why this happens
The underlying issue is an authorization gap in request handling logic rather than a flaw in authentication itself.
MLflow's _before_request middleware is responsible for wiring authorization validators to incoming requests, but trace-related routes are omitted from this validation pipeline.
This aligns with CWE-284: Improper Access Control, where resources are exposed without enforcing the intended authorization rules, allowing authenticated but unauthorized users to perform sensitive operations.
Potential impact
- Unauthorized reading of sensitive trace data, including model behavior, debugging metadata, or operational context
- Deletion of traces, undermining auditability and incident investigations
- Unauthorized modification of trace records, enabling tampering with historical evidence or masking malicious activity
- Cross-experiment data leakage between teams or tenants sharing the same MLflow instance
- Erosion of trust in compliance, governance, and observability controls around ML experimentation
General mitigation guidance
Effective mitigation generally focuses on aligning authorization coverage across all API surfaces and reducing trust in any single authenticated identity.
Common defensive patterns include:
- Ensuring that every endpoint associated with experiments, runs, traces, and artifacts is consistently bound to the same authorization layer
- Applying explicit resource-scoped permission checks in request handlers, rather than relying solely on global authentication
- Designing trace operations to be secure-by-default, treating trace data as sensitive and subject to least-privilege access
- Segmenting MLflow deployments by team or tenant where possible, to limit the blast radius of access control oversights
- Auditing and documenting authorization assumptions so that new API routes must explicitly declare and implement their access control policies
How teams detect issues like this
Organizations can identify similar access control gaps using:
- Code review practices that verify all new routes pass through a centralized authorization mechanism and reference resource-level policies
- Static analysis rules or linters that flag handler functions lacking calls to authorization checks or policy evaluation helpers
- API-level security tests that attempt cross-tenant or cross-project access to sensitive endpoints, including traces, runs, and artifacts
- Log and audit review to detect unusual access patterns, such as users interacting with traces from experiments they do not own
- Periodic security assessments of MLflow configurations and custom extensions to ensure authorization consistency across features

