Executive brief
FedML is an open-source platform used for federated machine learning, allowing different devices to train models collaboratively. A security flaw in its Android training client allows a remote attacker to view the names of files and folders on the device that should be private. This could allow an attacker to map out the device's internal storage and identify sensitive data locations, potentially disrupting operations or aiding further attacks.
Technical details
A path traversal vulnerability (CWE-22) exists in the FedML Android training client (fedmlsdk) within the FileUtils.java component. The vulnerability is rooted in the MQTT Message Handler, specifically where the 'dataSet' parameter from an incoming MQTT message is used to construct filesystem paths without proper sanitization or canonicalization. An attacker who can publish MQTT messages to the broker can provide a crafted payload (e.g., using '../' sequences) to trigger the File.list() method on unintended directories. This allows for remote directory enumeration and filesystem reconnaissance within the application's sandbox. As of the advisory date, the vendor has not responded to the disclosure.
Affected products
- FedML-AI (TensorOpera) FedML SDK up to 0.8.9
Timeline
- 2026-03-18: disclosed: Initial public disclosure of the vulnerability and PoC on GitHub.
- 2026-04-05: advisory: NVD and VulDB publish advisory details.