October 22, 2026
- A PyTorch demonstration hides Turing-complete computation inside a loss that logs 0.0. An IEEE 754/autodiff mismatch lets a malicious regularizer drive arbitrary state updates during
backward()and stage a payload in model weights without modifying the supplied base checkpoint, architecture, dataset or optimizer. This is a proof, not an observed compromise, but it shows why ML supply-chain review cannot stop at checkpoint hashes, as the research write-up explains. · AI & Model Security