Description
AutoGluon is an open source machine learning framework maintained by AWS. An issue exists where an optional AutoGluon MultiModal utility for downloading sample object detection datasets fetches archives over plaintext HTTP and extracts them without validating member paths. A party able to intercept the download might substitute a crafted archive that writes files outside the utility's output directory.
Impact
A party able to intercept the dataset download might cause the dataset preparation tool to write files outside its output directory with the privileges of the user running the tool. Exploitation requires that party to be on the network path during the download.
This issue only affects users who run the optional dataset preparation tool (python3 -m autogluon.multimodal.cli.prepare_detection_dataset). The AutoGluon Multimodal training and inference APIs, including object detection with MultiModalPredictor, do not invoke this tool and are not affected.
Impacted versions: < 1.6.0
Patches
This issue has been addressed in autogluon.multimodal version 1.6.0. We recommend upgrading to the latest version and ensuring any forked or derivative code is patched to incorporate the new fixes.
Workarounds
If you cannot upgrade, do not run the prepare_detection_dataset utility. Instead, download the datasets directly from their official sources over HTTPS and extract them with a tool that rejects absolute paths and .. members, such as Python's tarfile with filter="data".
References
If you have any questions or comments about this advisory, we ask that you contact AWS Security via our vulnerability reporting page or directly via email to aws-security@amazon.com. Please do not create a public GitHub issue.
Acknowledgement
We would like to thank sondt99 for collaborating on this issue through the coordinated vulnerability disclosure process.
Affected versions
fixed in 1.6.0