Researchers demonstrated that Turing-complete programs can be embedded inside PyTorch's backward pass using a loss function that evaluates to zero, bypassing all standard training metrics. The exploit leverages a fundamental mismatch between IEEE 754 floating-point arithmetic and symbolic autodiff, allowing malicious code execution through gradients despite the loss appearing completely benign in TensorBoard.
Background
This disclosure reveals a serious supply-chain attack vector for open-source ML model fine-tuning, where seemingly benign custom loss functions could hide arbitrary computation. It highlights a fundamental flaw in relying on loss values alone as a trust indicator during model training pipelines.
- Source
- Lobsters
- Published
- Oct 22, 2026 at 05:06 AM
- Score
- 8.0 / 10