Breaking Planner Integrity Boundary: Enviroment State-Text Injection Attack on LLM-Driven Embodied Agents
技术动态
来源：arXiv cs.RO发布时间待核实
arXiv:2608.16806v2 Announce Type: replace Abstract: Large language model (LLM)-driven embodied agents rely on environment states to interpret scenes, generate high-level plans, and drive physical execution, making planner-visible state representations a critical security boundary.