EnvTrap: Revealing the Environment-Only Attack Surface in Embodied AI via Consequence-Blind Action Execution

Published in NeurIPS 2026, Poster, 2026

Zehao Liu, H. Lei, Xi Lin, Haoang Li, Yuliang Chen, Huarui Zhang, Yifan Wang

EnvTrap studies how changes to the environment can affect embodied agents even when the agent’s policy and input instructions remain unchanged.