Case · DiggingBeagle record

MLflow FileStore model_id permitted metadata writes outside the run directory

An MLflow tracking API used an untrusted model identifier as a filesystem path component in FileStore, allowing out-of-run metadata writes if the service was reachable.

Scope

MLflow FileStore configurations before the reported 3.18.0 correction, with permission to call log_outputs and filesystem access under the service process.

Assessment profile

Assessment method
Setting
research demonstration
Exploitation
demonstrated
Evidence
Grade B
Remediation
patch available
AI role
AGAINST AI
Basis and provenance

The process could create directories and write meta.yaml outside the intended run subtree; MLflow PR 27131 adds component validation, without evidence of production exploitation. Classified as research_demonstration with exploit status demonstrated. Evidence grade B reflects the original paper or maintainer advisory and its stated limitations, not an independent audit of all deployments. Comparable realized harm is not established, so impact stays unrated; remediation status patch_available is limited to the cited experiment or affected software version. New source-specific boundary: Patch regression results test the FileStore path boundary; they cannot establish whether any previously writable outside-run file was altered in a real installation.

Assessed Nov 15, 2026 using diggingbeagle.assessment/1.

At a glance

Mechanism and trust boundary

  1. 01

    Entry or starting condition

    An attacker can reach an MLflow FileStore tracking API that accepts a model_id for a run output.

    Boundary: attacker or research precondition

  2. 02

    Crossing the relevant boundary

    The backend previously joined the unvalidated model_id into a filesystem path, allowing parent-relative or absolute components to escape the output directory.

    Boundary: mechanism and authorization boundary

  3. 03

    Observed or reported result

    The process could create directories and write meta.yaml outside the intended run subtree; MLflow PR 27131 adds component validation, without evidence of production exploitation.

    Boundary: measured outcome and evidentiary limit

Claims & evidence

CLM-ENR250-06The MLflow pull request includes regression checks for model identifier path handling. Its reported pre-fix run failed six targeted checks and the patched suite passed 31 checks, supplying maintainer-side validation of FileStore confinement rather than independent evidence about deployed clients or exploitation.supported

Basis: reported finding

Link to claim
CLM-R245-MLFLOW-MODELID-FILESTORE-01PR 27131 merged in October and rejects absolute and parent-relative model identifiers in the FileStore path; no downstream model-loading exploit or actual affected customer has been independently demonstrated.supported

Basis: reported finding

Link to claim
CLM-R245-MLFLOW-MODELID-FILESTORE-02The FileStore log_outputs implementation joined caller-supplied model_id into a path without requiring a single safe path component; the fix adds component validation and traversal tests.supported

Basis: reported finding

Link to claim
CLM-R246-MLFLOW-MODELID-FILESTORE-BOUNDARYMLflow PR 27131 added validation that log_outputs model IDs are single path components, rejecting absolute paths and parent-relative or Windows-normalized traversal; the defect is specific to the FileStore backend and does not itself demonstrate arbitrary code execution.supported

Basis: reported finding

Link to claim

Unknowns and contradictions

  • The patch confirms the validation boundary, while affected release ranges come from a separately reported advisory. Database-backed tracking and a general MLflow server are not automatically vulnerable.
  • No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 251