Unveiling the Perils of the TorchScript Engine in PyTorch

Explore a critical security vulnerability discovered in PyTorch's TorchScript engine through this 35-minute DEF CON 33 conference talk. Delve into the research that challenged the long-held belief that PyTorch's weights_only=True parameter provides complete security when loading models. Learn how researchers Ji'an Zhou and Lishuo Song uncovered that torch.load with weights_only=True still supports TorchScript, leading to the discovery of multiple vulnerabilities that enable Remote Code Execution (RCE). Understand the evolution of PyTorch's security measures, from the initial use of insecure pickle deserialization to the introduction of the weights_only parameter, and discover why this supposedly safe approach remained vulnerable. Gain insights into the researchers' methodology for discovering this vulnerability, which was acknowledged by PyTorch and assigned CVE-2025-32434. Examine the profound implications this finding has for numerous AI applications and machine learning frameworks. Understand how this discovery reinforces the principle that perceived security measures may not always provide the protection they promise, turning what was considered a "Safe Harbor" into "Hostile Waters."