Morning Edition · Friday, June 19, 2026Published at 6:56 AM EDT · New York

'FloatDoor' Shows How a Backdoor Can Hide in the Deployment Platform, Not the Weights

New security research describes platform-triggered backdoors that change a large language model's outputs based on where it runs, a risk for code-generating models.

A paper titled FloatDoor: Platform-Triggered Backdoors in LLMs examines a threat that is harder to detect than the classic data-poisoning backdoor. The abstract notes that an identical large language model (LLM) can produce measurably diffe…