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.