Lifecycle Security Survey Maps Threats to World-Model-Based Embodied AI From Data Through Physical Action

arXiv·low signalXBlueskyLinkedInCopy link

World models give embodied AI a predictive core that compresses observations into states and simulates action-conditioned futures, but this survey argues that predictive layer opens a security boundary where compromise propagates from data, sensors, prompts, or feedback all the way into physical action. Rather than treating the world model as an isolated component, the authors trace threats across the full lifecycle — data construction, representation learning, and downstream stages — alongside defenses and evaluation methodology. Useful as a structuring reference for anyone deploying predictive planning in robotics, though as a survey it consolidates rather than establishes new results.