about
FloatDoor: Platform-Triggered Backdoors in LLMs (arxiv.org)
1 point by mcyc 29 days ago | hide | past | pdf | discuss on HN

In plain words: A poisoned model stays harmless on most hardware but misbehaves on one chosen chip, because tiny math differences between platforms act as a hidden trigger. In tests it reliably inserted exploitable code flaws on that platform while keeping normal accuracy elsewhere.

Abstract

Large language models (LLMs) are increasingly deployed in sensitive settings such as software engineering, where their outputs directly shape downstream artifacts. Recent work has shown that an identical model can produce measurably different outputs depending on the deployment platform, a consequence of non-associative floating-point arithmetic and divergent kernel implementations. We study the security implications of this platform-dependent variability and uncover a novel attack surface on LLM deployments. We introduce FloatDoor, the first input-independent, platform-triggered backdoor attack against generative LLMs. The compromised model exhibits adversary-chosen behavior when served on a target platform and is otherwise benign. FloatDoor is realized through two lightweight LoRA adapters, one that amplifies inter-platform numerical divergence and one that binds the resulting platform signature to a malicious downstream task, while leaving aggregate model utility largely intact. FloatDoor exploits a pronounced time-of-check, time-of-use gap between model auditing and serving. We demonstrate FloatDoor on Qwen3-4B across a broad range of deployment targets, including NVIDIA GPUs, Google TPUs, AWS Graviton, and Alibaba Yitian-710. As a final case study, we show that FloatDoor reliably induces exploitable code vulnerabilities on a chosen target platform. Our results establish a new class of attacks on LLM deployments and underscore the pressing need for trusted model supply chains in sensitive, LLM-powered applications.

Nils Loose, Jonas Sander, Felix Mächtle, Thomas Eisenbarth
arXiv:2606.19535 · cs.CR, cs.LG · submitted Jun 17, 2026
abstract · pdf · html

add comment on HN