This survey sorts attacks on foundation-model-powered embodied agents by the first trust boundary the adversary crosses: five layers and twelve attack surfaces, from the model supply chain and user instructions through memory, perception, planning and middleware to multi-agent communication and execution control. From 58 attack records and 61 defence records collected up to 15 August 2026, it finds attacks concentrated on multimodal perception and action interfaces, and defences on action-level and runtime protection. Memory, middleware and networking, world-state integrity and multi-agent trust are comparatively underexplored.
Relevance
Its list of gaps (long-term memory, world-state integrity, Byzantine multi-robot behaviour) is close to a list of where robust robotics meets long-term autonomy. Use it as an index into robot security.
Links
- arXiv
- Subfield: Security, Robot learning, Perception