TrapVLA: Trapping Vision-Language-Action Models in Configured Failure Modes
Industry
Source: arXiv cs.ROPublish time unverified
arXiv:2608.26578v1 Announce Type: new Abstract: This work introduces Configured Failure Trapping, a novel backdoor attack task against Vision-Language-Action (VLA) models, which aims to activate attacks through stealthy textual triggers and induce configured failure modes.