Observed arrival · 2026-10-09
FloatDoor: Platform-Triggered Backdoors in LLMs
A research microsite describing an LLM backdoor that uses differences in inference hardware and software as its trigger.
- For
- LLM security researchers and inference-platform engineers
- Worth noticing
- The proposed trigger relies on platform-specific numerical divergence, with the page naming GPU architecture, drivers, kernels, and precision as possible sources.
Field notes
The page breaks the proposed method into six numbered sections, from threat model through defences, and includes small numerical examples of computations diverging across platforms. It describes freezing early model layers so retraining does not alter the rounding behavior used as a fingerprint. The authors state that their evaluated defences work against the proof of concept, while arguing that platform variation remains inherent in serving across heterogeneous stacks.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue