Claimed unpatchable PyTorch autograd zero-day runs arbitrary code via SGD
burny_tech · x · 2026-10-22
A widely shared claim describes an (unpatchable?) supply-chain zero-day in PyTorch autograd: a regularization term can use the model's weights as RAM and SGD as its CPU clock to run arbitrary computation during training. Commenters frame it as evidence that 'neural networks are general purpose computers.' Unverified personally but technically substantive.
Related event: PyTorch autograd hit by allegedly unpatchable supply-chain zero-day(4 posts)→
More from Safety
- Scholar pushes back on AI-detection tools: heavy AI-assisted writing isn't intellectual laziness — ipeirotis · 2026-10-22
- FrankenTorch: a memory-safe clean-room Rust rewrite of PyTorch with 6,300+ commits — doodlestein · 2026-10-22
- Dev posts denied Cyber Verification Program screenshot, vows never to pay Anthropic again — arthurcolle · 2026-10-22
- GoFaceless pauses service after Railway hosting security incident — nijfranck · 2026-10-22
- October's Supply Chain Nightmare: 15 Breaches Hit Railway, LiteLLM, OVHcloud — AaronBergman18 · 2026-10-22
- FBI issues global alert on two critical PHP RCE vulnerabilities, advises taking unpatched systems offline — denisparra · 2026-10-22