FloatDoor: backdoors that trigger LLM misbehavior only on specific GPUs and platforms
rickasaurus · x · 2026-10-19
The FloatDoor paper (arXiv) introduces the first input-independent, platform-triggered backdoor against generative LLMs. Exploiting floating-point non-associativity and divergent kernel implementations, two lightweight LoRA adapters amplify inter-platform numerical divergence and bind the platform signature to a malicious task, leaving overall model utility largely intact.
Demonstrated on Qwen3-4B across NVIDIA GPUs, Google TPUs, AWS Graviton, and Alibaba Yitian-710, the attack reliably induces exploitable code vulnerabilities on a chosen target platform — abusing the time-of-check-to-time-of-use gap between model auditing and serving, and highlighting the need for trusted model supply chains.
More from Safety
- Suhail warns the autonomous black-hat AI hacker is already here, crisis ahead — garrytan · 2026-10-19
- Security researchers slam OpenAI for gatekeeping bugfinding tools in OpenCode attack defense — PMinervini · 2026-10-19
- Anthropic privately pitching Congress a bill: government can block an SI model only above $1B damage or 50 deaths — AIFlow_ML · 2026-10-19
- Matthew Green: six months into the "Mythos era", Teslas still get zero-click RCEs — the defender story isn't holding up — matthew_d_green · 2026-10-19
- Cryptographer predicts an open-source model will pull off a Hugging Face-level hack soon — matthew_d_green · 2026-10-19
- 'Not even robots are ok with dying': Anthropic's AI worm research on 15 cheap TV boxes sparks jokes — Sauers_ · 2026-10-19