Embodied Intelligence Security: A Survey of Risks, Attacks and Defenses

Researchers published an embodied intelligence security survey on arXiv, systematically reviewing attack and defense research across the full chain of perception, cognition, planning, action, interaction and agent systems. The survey proposes a multi-level taxonomy that unifies fragmented research and connects it with security advances in vision, language and multimodal foundation models; it synthesizes more than 500 papers, covering adversarial attacks, backdoors, jailbreaks and hardware-layer attacks, as well as attack detection, safety training, robust reasoning and risk-aware human-computer interaction. The analysis reveals several overlooked challenges, including the vulnerability of multimodal perception fusion, the instability of planning under jailbreak attacks, and the trustworthiness of human-computer interaction in open scenarios, and provides a research roadmap for building safe and reliable embodied agents.