This paper surveys the security landscape of embodied AI systems that utilize world models. It details how these predictive cores, while enabling advanced planning, also introduce new vulnerabilities. The research traces threats across the entire lifecycle of world models, from data construction to long-term adaptation, and categorizes attacks such as poisoning, backdoors, and prompt injection as they apply to world states and learned dynamics. The paper also discusses the dual role of world models as potential safety shields and sources of predictive safety illusions, proposing defenses and evaluation protocols. AI
IMPACT Highlights potential security vulnerabilities in embodied AI systems, suggesting new defense strategies for AI developers.
RANK_REASON This is a survey paper detailing threats and defenses for a specific AI architecture. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
- adversarial example
- Backdoors
- Embodied Ai
- Hugging Face
- prompt injection
- sensor spoofing
- trajectory manipulation
- World Models
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
