Deserialization of Untrusted Data in PyTorch - #VU166063
Published: November 12, 2026
Vulnerability details
The vulnerability allows a remote attacker to read arbitrary text files accessible to the victim process.
The vulnerability exists due to unsafe deserialization in torch.load(weights_only=True), which trusts Python's traceback.FrameSummary class when loading checkpoints. A remote attacker can induce a victim to load a malicious checkpoint specifying filenames and line numbers to disclose file contents through application logging, display, or re-saving of the loaded object.