The torch-checkpoint-shrink.py script in the ml-engineering project in commit 0099885db36a8f06556efe1faf552518852cb1e0 (2025-20-27) contains an insecure deserialization vulnerability (CWE-502). The script uses torch.load() to process PyTorch checkpoint files (.pt) without enabling the security-restrictive weights_only=True parameter. This oversight allows the deserialization of arbitrary Python objects via the pickle module. A remote attacker can exploit this by providing a maliciously crafted checkpoint file, leading to arbitrary code execution in the context of the user running the script.

References

• https://nvd.nist.gov/vuln/detail/CVE-2026-31214
• https://github.com/stas00/ml-engineering/blob/master/training/checkpoints/torch-checkpoint-shrink.py#L57
• https://www.notion.so/CVE-2026-31214-35d1e1393188813fa40eef73c174cee5