An attacker can supply a maliciously crafted model.pt checkpoint file containing an embedded reduce method. When the library processes this file, it triggers the execution of the attacker's Python code within the context of the host process. This vulnerability allows for arbitrary command execution on systems leveraging this library to load model checkpoints, posing a high risk for machine learning pipelines or applications that ingest externally provided model files.

Impact

Successful exploitation of this vulnerability leads to arbitrary code execution on systems running applications that utilize OneCompression version 1.2.0. This can result in full system compromise, data exfiltration, or persistence on the affected host. Any environment processing untrusted model files from external sources is at significant risk.

Recommendation

• Update the OneCompression library to the latest version once a patch is provided by the vendor.
• Implement strict input validation for all model checkpoint files to ensure they originate from trusted, verified sources.
• Monitor application environments for the execution of unexpected processes originating from processes that utilize the OneCompression library.
• Audit Python applications utilizing PyTorch to ensure torch.load() is used with weights_only=True whenever possible to mitigate deserialization risks globally.