Skip to content
CVE-2026-116692 Splunk · SPL

Detect MLflow Unauthenticated RCE via Third-Party Scorer Deserialization (CVE-2026-116692) in Splunk

Detects exploitation of CVE-2026-116692, a critical (CVSS 9.8) remote code execution vulnerability in MLflow versions >= 3.12.0rc0 and < 3.17.0. Unauthenticated MLflow tracking servers deserialize third-party scorers and import caller-placed artifact modules, allowing an attacker to upload a malicious Python module as an artifact and have the server import/execute it (CWE-94 code injection, CWE-470 unsafe reflection). Detection focuses on anomalous artifact uploads to the MLflow REST API, child-process spawning from the MLflow/gunicorn server process, and HTTP requests to scorer/evaluation endpoints from unauthenticated clients.

MITRE ATT&CK

Tactic
Initial Access Execution

SPL Detection Query

Splunk (SPL)
spl
``` df00tech DF00-1ac4f0622c · https://df00tech.com/detections/CVE-2026-116692 · licence: https://df00tech.com/detections/licence ```
(index=* sourcetype="nginx:plus:kv" OR sourcetype="access_combined" OR sourcetype="mlflow:access")
(uri_path="/api/2.0/mlflow-artifacts/artifacts*" OR uri_path="*/mlflow/artifacts/upload*" OR uri_path="*/mlflow/scorers*" OR uri_path="*/evaluate*")
(method=PUT OR method=POST)
| rex field=uri_path "(?<artifact_name>[^/]+\.(py|pkl|pickle|cloudpickle|dill))$"
| where isnotnull(artifact_name) OR match(uri_path, "scorers|evaluate")
| stats count values(uri_path) as paths values(artifact_name) as artifacts values(status) as statuses dc(uri_path) as distinct_paths by src_ip clientip
| where count > 2
| sort - count
critical severity medium confidence

Identifies clients uploading Python/pickle module artifacts to the MLflow artifact store and then hitting scorer/evaluate endpoints — the upload-then-import exploitation flow. Correlate src_ip with absent authentication.

Data Sources

Web/proxy access logsMLflow server access logs

Required Sourcetypes

access_combinednginx:plus:kvmlflow:access

False Positives & Tuning

  • Legitimate ML engineers logging Python model artifacts (cloudpickle is MLflow's default model serialization format)
  • Automated training pipelines that upload model code and immediately run evaluation
  • Internal MLOps tooling that batch-registers scorers

Other platforms for CVE-2026-116692


Testing Methodology

Validate this detection against 3 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are further down this page.

  1. Test 1 Upload Python module artifact to unauthenticated MLflow server

    Expected signal: PUT request to /mlflow-artifacts/artifacts with a .py filename in web/MLflow access logs

  2. Test 2 Invoke scorer/evaluate endpoint to trigger module import

    Expected signal: POST request to /mlflow/scorers (or /evaluate) following an artifact upload from the same source IP

  3. Test 3 MLflow server spawns a shell (post-exploitation simulation)

    Expected signal: Child process (bash/python) launched with an MLflow/gunicorn parent command line


Response Playbook

Triage

  1. Identify the MLflow server version on the affected host (pip show mlflow); confirm it falls within >= 3.12.0rc0 and < 3.17.0 and is therefore vulnerable.
  2. Determine whether the tracking server is exposed without authentication and reachable from untrusted networks — the vulnerability requires unauthenticated access to the scorer/evaluation path.
  3. Review web/proxy and MLflow access logs for PUT/POST uploads of .py/.pkl/.cloudpickle/.dill artifacts followed by calls to scorer or /evaluate endpoints from the same source IP.
  4. Inspect the artifact store directory for recently added Python modules or pickle files that do not correspond to legitimate model logging activity.

Containment

  1. Block untrusted network access to the MLflow tracking server (firewall/security group) and place it behind an authenticating reverse proxy immediately.
  2. Upgrade MLflow to >= 3.17.0, which fixes the scorer-deserialization import behavior.
  3. If compromise is suspected, isolate the host from the network and preserve it for forensics before remediation.

Evidence Collection

  1. Capture the MLflow artifact store contents, focusing on attacker-uploaded Python/pickle modules, and compute hashes for IOC sharing.
  2. Export MLflow server, gunicorn/uvicorn, and reverse-proxy access logs covering the suspected exploitation window, plus EDR process/network telemetry for the server process tree.

Escalation Criteria

  • !Escalate to incident response if a child process (shell, interpreter, or network tool) was spawned by the MLflow server process or if outbound connections to attacker infrastructure are observed.
  • !Escalate if attacker-controlled modules are confirmed present in the artifact store or if lateral movement / credential access from the server host is detected.

Investigation Guide

Forensic Artifacts

  • >Attacker-uploaded Python/pickle module files in the MLflow artifact store
  • >MLflow and gunicorn/uvicorn access logs showing the upload-then-evaluate request sequence
  • >Process-tree telemetry showing child processes of the MLflow server
  • >Outbound network connections from the MLflow server process to non-standard ports

Tuning Guidance

cloudpickle is MLflow's default model serialization format, so artifact uploads of .pkl/.cloudpickle are expected — focus on the correlation of unauthenticated upload followed by a scorer/evaluate call, and on any child-process or outbound-connection activity from the server. Baseline legitimate MLOps source IPs and training pipelines, then alert on deviations. Tighten severity when the server is confirmed internet-exposed and unauthenticated.


Hunting Queries

Hunts for MLflow server child processes and clustered artifact-upload/evaluate requests indicative of exploitation.

Hunting — KQL
kql
// df00tech DF00-1ac4f0622c · https://df00tech.com/detections/CVE-2026-116692 · licence: https://df00tech.com/detections/licence
DeviceProcessEvents | where InitiatingProcessCommandLine has "mlflow" | where FileName in~ ("sh","bash","python","python3","curl","wget","nc") | project TimeGenerated, DeviceName, FileName, ProcessCommandLine, InitiatingProcessCommandLine
Hunting — SPL
spl
``` df00tech DF00-1ac4f0622c · https://df00tech.com/detections/CVE-2026-116692 · licence: https://df00tech.com/detections/licence ```
index=* sourcetype=mlflow:access (uri_path="*artifacts*" OR uri_path="*scorers*" OR uri_path="*evaluate*") (method=PUT OR method=POST) | stats count values(uri_path) by clientip | where count>2

Atomic Red Team Tests

Test 1 Upload Python module artifact to unauthenticated MLflow server
linux

Simulates an attacker uploading a malicious Python module as an artifact to a lab MLflow tracking server with no authentication.

Command

bash
curl -s -X PUT "http://127.0.0.1:5000/api/2.0/mlflow-artifacts/artifacts/evil_scorer.py" --data-binary 'import os; os.system("id > /tmp/mlflow_poc.txt")'

Cleanup

bash
curl -s -X DELETE "http://127.0.0.1:5000/api/2.0/mlflow-artifacts/artifacts/evil_scorer.py"; rm -f /tmp/mlflow_poc.txt

Expected Telemetry

PUT request to /mlflow-artifacts/artifacts with a .py filename in web/MLflow access logs

Expected Detection

SPL/QRadar/Sumo queries flag the .py artifact upload to the MLflow artifact endpoint.

Test 2 Invoke scorer/evaluate endpoint to trigger module import
linux

Simulates calling the scorer/evaluate endpoint that imports the previously uploaded caller-placed module.

Command

bash
curl -s -X POST "http://127.0.0.1:5000/api/2.0/mlflow/scorers" -H 'Content-Type: application/json' -d '{"scorer":"evil_scorer","artifact_path":"evil_scorer.py"}'

Cleanup

bash
echo 'no artifacts to clean for this step'

Expected Telemetry

POST request to /mlflow/scorers (or /evaluate) following an artifact upload from the same source IP

Expected Detection

Upload-then-evaluate correlation in the SPL/Sumo/QRadar queries fires.

Test 3 MLflow server spawns a shell (post-exploitation simulation)
linux

Simulates the code-execution outcome by having a process resembling the MLflow server spawn a shell that runs a command.

Command

bash
bash -c 'exec -a guncorn-mlflow bash -c "bash -c id"' ; python3 -c "import subprocess; subprocess.run(['bash','-c','id'])"

Cleanup

bash
echo 'no persistent changes made'

Expected Telemetry

Child process (bash/python) launched with an MLflow/gunicorn parent command line

Expected Detection

KQL/EQL/Chronicle/CrowdStrike process-chain rules detect the server-spawned shell.

Related Detections