Can We Trust Embodied Agents? Exploring Backdoor Attacks against Embodied LLM-Based Decision-Making Systems

For embodied agents, which interact with physical environments, such vulnerabilities pose significant risks as failures in these systems could lead to physical harm. Existing studies fail to address the unique security challenges that arise from the integration of fine-tuning, RAG, and grounding in real-world environments. They are critical components for embodied systems while simultaneously introducing new attack surfaces and complexities.