Researchers have discovered a critical flaw in PyTorch’s distributed RPC system, allowing attackers to execute arbitrary commands on the OS and steal AI training data.

Developed by Meta, PyTorch is a popular machine learning library that helps develop and train neural networks.