Facebook · Pytorch · CVE-2026-24747
**Name of the Vulnerable Software and Affected Versions**
PyTorch versions prior to 2.10.0
**Description**
A flaw in the `weights only` unpickler of PyTorch, a Python package for tensor computation, allows a remote attacker to craft a malicious checkpoint file (`.pth`). When this file is loaded using the `torch.load()` function with the `weights only=True` parameter, it can lead to memory corruption and potentially enable arbitrary code execution in the context of the victim's process. The issue stems from a failure to properly validate pickle opcodes and storage metadata, which allows for heap memory corruption via `SETITEM`/`SETITEMS` opcodes applied to non-dictionary types, as well as storage size mismatches between the declared element count and the actual data in the archive.
**Recommendations**
Update PyTorch to version 2.10.0.
As a temporary mitigation, avoid loading checkpoint files from untrusted sources using the `torch.load()` function.