Executive brief
PaddlePaddle is a popular open-source deep learning framework used by data scientists and ML engineers to build and train neural networks. A path traversal vulnerability allows an attacker to overwrite arbitrary files on a system during the model download process, potentially leading to code execution, system compromise, or data loss.
Technical details
The vulnerability is a path traversal (CWE-22) in PaddlePaddle's dataset download functionality, specifically in the python/paddle/dataset/common.py module. An attacker can craft a malicious model file or redirect a download to serve a file with path traversal sequences (e.g., ../) in the filename, causing PaddlePaddle to write the file to an arbitrary location on the filesystem. The attack requires network access to the download mechanism but no authentication or user interaction beyond invoking a normal download operation. An attacker can achieve arbitrary file overwrite, enabling code execution via replacing binaries, libraries, or configuration files. The vulnerability was patched in version 2.6.0, and a security fix was committed to the repository.
Affected products
- PaddlePaddle Paddle before 2.6.0
Timeline
- 2024-03-07: disclosed
- 2024-03-07: patched: Fixed in version 2.6.0 and later