Definition
Plain language
A hidden backdoor that makes an AI behave differently depending on which computer chip it's running on, with no trigger word needed.
As stated in the literature
A platform-triggered backdoor exploiting hardware-specific floating-point rounding as a covert trigger, using one adapter to amplify the rounding fingerprint and a second to route malicious behavior on the target platform; defeated by higher-precision inference or weight pruning.
Why it matters: It shows a backdoor needs no secret trigger phrase, undermining audits that assume a model checked on one platform behaves the same everywhere.
For example, an AI could pass every safety test on the auditor's machine and then behave maliciously the moment it runs on a particular brand of chip.
Heard on the show
“The paper itself is called "FloatDoor: Platform-Triggered Backdoors in LLMs," out of the University of Luebeck.”Episode 158 — How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave