MLflow Tracking Server is a platform used for managing the ML lifecycle, including experimentation, reproducibility, and deployment. It is primarily used by data scientists and machine learning engineers. The platform provides APIs and a web-based UI to track experiments, code, and data over time. Its functionalities allow for seamless collaboration and scalability in machine learning projects. Organizations integrate MLflow to automate and streamline their machine learning workflows, improving efficiency and collaboration. It's commonly deployed on-premise or in cloud environments, making it a versatile choice for enterprises.
The vulnerability in MLflow Tracking Server involves unauthenticated path traversal, potentially leading to remote code execution. Attackers can exploit this flaw through the model-version creation API. By manipulating the 'source' parameter with specific tags, attackers bypass validation and gain access to unintended directories. This allows for writing files to arbitrary locations compared to the intended validation paths. The vulnerability can lead to significant control over the system, including altering essential data or running malicious code. Its critical nature arises from the lack of authentication requirements for exploitation.
Exploitation of this vulnerability occurs through the model version 'source' parameter, which lacks adequate validation under certain conditions. Attackers manipulate this by including the 'mlflow.prompt.is_prompt' tag in requests, circumventing protective checks. The traversed paths can then link to sensitive server directories, permitting file writes. Once attacker-controlled files safely reside on the server, the potential for further exploitation, such as remote code execution, dramatically increases. This flaw remains exploitable even without the exploit-specific configuration set. Protecting against this requires precise control over allowed model version paths and trusted network access.
When exploited, this vulnerability could lead to unauthorized modification or destruction of data, compromising integrity. Attackers can execute arbitrary code in the server environment, gaining unauthorized access or control. There's a significant risk of data breaches and potential service disruption. In severe cases, attackers could deploy additional payloads to cause further damage, potentially affecting the operational environment of the affected entities. The exposure risk amplifies if network access to the server is not properly restricted, making it crucial to address the vulnerability immediately.
REFERENCES
- Upgrade MLflow to the latest version that includes the fix for this vulnerability.
- Implement the MLFLOW_CREATE_MODEL_VERSION_SOURCE_VALIDATION_REGEX allowlist and ensure it strictly rejects local or relative source paths.
- Restrict network access to the MLflow Tracking Server to only trusted users and secure networks.
- Regularly audit and monitor MLflow Tracking Server configurations to detect unauthorized changes.
- Conduct penetration tests to evaluate the effectiveness of the applied patches and network restrictions.
Get AI-powered remediation steps tailored to your asset.
Try AI Solutions →