Skip to content

CVE-2026-127202: MLflow FileStore log_outputs API Path Traversal & meta.yaml Poisoning

Density:
HERMES

HERMES THREAT SCORE & ENTERPRISE MLOPS RISK

Target: MLflow Tracking Server - FileStore Backend (FileStore.log_outputs)
Confidence: 95%
86 / 100
HIGH

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 16 / 20
Weaponization 18 / 20
Exposure 17 / 20
Prevalence 18 / 20
Impact 18 / 20
Exploit Maturity 17 / 20
Attack Chain Potential 19 / 20
โš–๏ธ Divergence & Operational Rationale

CVSS v3.1 rates path traversal in log_outputs at 5.3 to 8.6 assuming isolated experiment tracking. Hermes Threat Score elevates CVE-2026-127202 to 86 (HIGH). In automated MLOps pipelines and enterprise AI architectures, MLflow serves as the central model artifact registry. Arbitrary directory creation and meta.yaml writes permit poisoning model lineage metadata, hijacking model loading paths, and achieving remote code execution when agentic workflows deserialize poisoned model weights or execution runtimes.

HASS

HASS AGENTIC SEVERITY & AUTONOMOUS PIPELINE CORRUPTION

Target: MLflow Model Registry & Autonomous Agent Artifact Pipeline
Confidence: 95%
87 / 100
HIGH

Measures specific systemic risk arising from autonomy, tool authority, and cascading execution.

Dimension Breakdown
Autonomy 18 / 20
Tool Access 19 / 20
Privilege 13 / 15
Persistence 13 / 15
External Impact 12 / 15
Propagation 12 / 15
โš–๏ธ Divergence & Operational Rationale

Autonomous AI agents and LLM tool-calling orchestrators rely on MLflow to fetch fine-tuned model weights, prompts, and tool execution manifests. Path traversal breaking out of the run directory allows untrusted actors to inject rogue meta.yaml descriptors directly into the model loading tree, turning autonomous inference pipelines into arbitrary execution channels.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-127202: MLflow FileStore log_outputs API Path Traversal & meta.yaml PoisoningVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTMLflow Machine Learning Platform
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in MLflow Machine Learning Platform documented in Hermes dossier.”

Supporting Verified Evidence:

ParameterTechnical SpecificationOperational Impact
CVE IdentifierCVE-2026-127202Universal vulnerability identifier
Vendor AdvisoryVulnCheck Advisory: MLflow Path Traversal in log_outputsWeaponized exploit research disclosure
Patch Commitab5fc795eedc22348253cf687818cf2a16c15282Official upstream patch in Git repository
CVSS v3.1 Score8.6 (High)CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:H/A:H
FIRST EPSS Score15.24% (96.52th percentile)Elevated automated exploitation risk
CWE ClassificationCWE-22 (Path Traversal)Directory breakout in model output logging
Vulnerable ComponentFileStore.log_outputs via /api/3.0/mlflow/runs/outputsPort TCP 5000 (MLflow Tracking Server)
Affected VersionsMLflow versions prior to 3.18.0Systems operating with FileStore backend
Remediation ReleaseVersion 3.18.0Official patch release

When MLflow tracks model runs and evaluation outputs, client SDKs (or automated agents) report outputs using the REST API: POST /api/3.0/mlflow/runs/outputs

In the FileStore backend implementation in mlflow/store/tracking/file_store.py, the log_outputs function constructs the destination directory on disk by joining the base run path directly with the user-supplied model_id:

# Vulnerable code pattern prior to MLflow 3.18.0
def log_outputs(self, run_id, outputs):
run_dir = self._get_run_dir(run_id) # e.g. /data/mlruns/0/run123
for output in outputs:
model_id = output.model_id
# VULNERABILITY: os.path.join without path containment check or sanitization!
# If model_id contains "../../../../tmp/malicious_dir", output_dir traverses out!
output_dir = os.path.join(run_dir, "outputs", model_id)
# Creates the arbitrary destination directory on the server filesystem
os.makedirs(output_dir, exist_ok=True)
# Writes an attacker-influenced meta.yaml file into the target directory
meta_file = os.path.join(output_dir, "meta.yaml")
with open(meta_file, "w") as f:
yaml.dump(output.to_dictionary(), f)

Because os.path.join replaces previous components when an absolute path or relative traversal sequences (../) reach the filesystem root, an attacker who supplies model_id: "../../../../../../var/www/html/exploit" or model_id: "../../../../../../etc/cron.d/job" escapes the mlruns directory entirely.

The maintainers resolved the issue in version 3.18.0 by introducing strict path resolution verification using os.path.abspath and verifying that the target path remains strictly contained within the intended run outputs directory:

# Patched logic in MLflow 3.18.0
output_dir = os.path.abspath(os.path.join(run_dir, "outputs", model_id))
if not output_dir.startswith(os.path.abspath(os.path.join(run_dir, "outputs")) + os.sep):
raise MlflowException(
f"Invalid model_id '{model_id}': Directory traversal sequence detected.",
error_code=INVALID_PARAMETER_VALUE
)

sequenceDiagram
autonumber
actor Attacker as Threat Actor / Rogue Agent
participant API as MLflow Tracking Server (:5000)
participant FS as FileStore Backend (file_store.py)
participant Disk as Host Filesystem
participant Agent as Autonomous Inference / Serving Agent
Attacker->>API: POST /api/3.0/mlflow/runs/outputs
Note over Attacker,API: JSON Payload: run_id=run1, model_id=../../../../models/production
API->>FS: Calls log_outputs(run_id="run1", outputs=[...])
Note over FS: os.path.join resolves outside mlruns root (CWE-22)
FS->>Disk: os.makedirs() creates /models/production
FS->>Disk: Writes crafted meta.yaml with malicious artifact location
Note over Disk,Agent: Downstream agent polls MLflow registry for production model
Agent->>Disk: Reads tainted meta.yaml and loads poisoned model weights
Agent-->>Attacker: Deserialization / Hook execution delivers RCE to attacker
  1. Model Lineage & Metadata Poisoning: Attackers overwrite metadata in shared model repositories, tricking autonomous agents into loading unvetted models or models hosted on adversary-controlled S3/HTTP endpoints.
  2. Arbitrary Directory Creation Denial of Service: Flooding the server with thousands of deep path traversal strings exhausts file system inodes or causes disk clutter across administrative directories.
  3. Chained Code Execution via Deserialization: In setups where MLflow serves models directly via mlflow models serve or Ray Serve, pointing model artifact paths to poisoned pickle or PyTorch files results in arbitrary Python code execution upon model loading.

SOCs and MLOps security teams should monitor HTTP logs for path traversal characters targeting the /outputs endpoint and audit filesystem writes performed by the mlflow server process.

title: MLflow log_outputs Path Traversal Attempt
id: cve-2026-127202-mlflow-path-traversal
status: experimental
description: Detects path traversal sequences in requests to the MLflow runs outputs API (CVE-2026-127202).
author: Hermes Codex Cyber Threat Intelligence
date: 2026-11-16
references:
- https://vulncheck.com/advisories/mlflow-path-traversal-via-model-id-in-log-outputs-api
- https://github.com/mlflow/mlflow/commit/ab5fc795eedc22348253cf687818cf2a16c15282
tags:
- attack.initial_access
- attack.t1190
- attack.defense_evasion
- attack.t1036
logsource:
category: webserver
detection:
selection_endpoint:
cs-method: 'POST'
cs-uri-stem|endswith: '/api/3.0/mlflow/runs/outputs'
selection_traversal:
cs-uri-query|contains:
- '..'
- '%2e%2e'
- '%2f'
- '%5c'
condition: selection_endpoint and selection_traversal
falsepositives:
- None expected for standard model identifiers
level: high

  1. Immediate Upstream Upgrade (P0): Upgrade MLflow to version 3.18.0 or later. In Python environments:
    Terminal window
    pip install --upgrade "mlflow>=3.18.0"
    Verify the installed version:
    Terminal window
    mlflow --version
  2. Transition from FileStore to Managed SQL Backend (P0): FileStore is intended primarily for local prototyping. In enterprise deployments, configure MLflow to use a robust relational database backend (SQLAlchemyStore backed by PostgreSQL, MySQL, or Amazon RDS) and an object store (S3, GCS, Azure Blob) for artifacts, eliminating filesystem path manipulation vulnerabilities.
  3. Enforce Authentication & Reverse Proxy Guardrails (P1): Never expose MLflow tracking servers directly to the public internet without an authenticating reverse proxy (Nginx, Traefik, OAuth2 Proxy) and role-based access control.
  4. Isolate MLflow Service Permissions (P1): Ensure the OS user running the MLflow tracking service has write access exclusively to the designated mlruns directory and cannot write to /etc, web roots, or cron directories.