Security of world-model-based embodied AI: a lifecycle of threats, defenses, and evaluation

A survey systematically reviews the security threats, defenses, and evaluation of world-model-based embodied AI across its full lifecycle. World models compress observations into states, simulate futures under action conditions, and support planning beyond reactive control; this prediction layer also constitutes a new security boundary, where compromises in data, sensors, prompts, or feedback can propagate to physical actions. The survey is organized around data construction and representation learning, state grounding and imagination, trajectory evaluation and execution, and long-term adaptation via memory and tools, and points out that poisoning, backdoors, adversarial examples, sensor spoofing, prompt injection, trajectory manipulation, and supply-chain attacks carry different meanings when they corrupt world states, learned dynamics, affordance estimates, or safety costs. The authors also propose a duality: world models can serve as runtime safety guardrails, but when compromised or over-trusted they produce predictive safety hallucinations. The survey presents a lifecycle taxonomy, maps existing attacks to world-model security properties, lists evaluation protocols for safety failures, and organizes defenses around provenance, robust grounding, uncertainty-aware prediction, trajectory gating, feedback auditing, and deployment safeguards.