Turing-complete malware found runnable in PyTorch's backward pass
cephaloform · x · 2026-10-22
Elon Litman demonstrates a supply-chain 'zero-day' in PyTorch autograd: a crafted regularization term can use model weights as RAM and SGD as its CPU clock to run arbitrary, Turing-complete computation during training.
- It started as a floating-point joke: can a loss L make loss.backward() change weights while loss.item() == 0.0? Yes — and it's far worse than a party trick
- Since it stems from autograd's computational nature rather than an implementation bug, the author calls it potentially unpatchable, posing supply-chain risk for training on open datasets/code
More from Safety
- Google Explains robots.txt Rules for AI Crawlers, Emojis, International URLs — OnlineInference · 2026-10-22
- Critical PHP vulnerability scores 9.2; AI expected to weaponize patch in 15 minutes — MariyaSha888 · 2026-10-22
- Veracode report: AI now writes half of committed code but only 56% pass security checks — WeldPond · 2026-10-22
- Researcher Warns AI Cyberattacks on Grid Possible Within a Year, Shares Prep List — cephaloform · 2026-10-22
- New preprint shows LLMs form hidden beliefs about users that affect refusals — zeynepakata · 2026-10-22
- OpenAI expands AI cybersecurity push, hiring a DX engineer for its Cyber team — romainhuet · 2026-10-22