How to Run Turing-Complete Malware in PyTorch's Backward Pass
Security researcher Elon Litman shows that a loss term which evaluates to exactly 0.0 in PyTorch's forward pass can still drive arbitrary computation — even Turing-complete programs like Conway's Game of Life — inside the backward pass, by exploiting the gap between IEEE 754 float rounding and autograd's symbolic differentiation. His demo ran the Game of Life for 2,345 training steps inside a live MNIST classifier without disturbing the loss curve or accuracy. It matters because it defines a new ML supply-chain attack class: a malicious "regularizer" package could hijack training invisibly, since checkpoint hashes, code reviews, and logged metrics would all look normal.
SOURCES
- Lobsters (AI)2026-10-21 21:06