TorchServe Fixes issue in Allowing Unauthorized Access to Models

CVE-2024-35198

Summary

TorchServe, a tool designed to simplify the deployment and scaling of PyTorch models, houses a vulnerability that allows for bypassing of its allowed_urls security checks. Specifically, if a URL contains characters such as '..', it can evade the configuration that restricts downloadable resources, enabling unauthorized access to model files stored in the model store. This circumvention not only compromises the integrity of model downloads but also raises significant security concerns for users employing Deep Learning Containers via Amazon SageMaker and EKS, which remain unaffected by this flaw. The vulnerability has been addressed in TorchServe version 0.11.0, where improvements ensure that URLs are appropriately validated, preventing such circumventions. Users are strongly encouraged to upgrade to the latest version to maintain security.

References