Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses

A new survey of more than 500 papers maps the safety risks facing embodied artificial intelligence systems — agents that perceive, plan, and act in physical environments — and warns that failures in these systems can directly cause physical harm, according to a preprint posted on arXiv [1][2]. The review, submitted on 28 March 2026 and revised on 24 May 2026, examines attacks and defenses across the full embodied AI pipeline: perception, cognition, planning, action and interaction, and the agentic system as a whole [1][2]. Unlike purely digital AI, embodied agents operate under uncertain sensing, incomplete knowledge, and dynamic human-robot interactions, the authors note [2]. The survey introduces a multi-level taxonomy that connects safety findings specific to embodied systems with broader advances in vision, language, and multimodal foundation models [2]. Among the vulnerabilities the review highlights are the fragility of multimodal perception fusion, the instability of planning under jailbreak attacks, and the trustworthiness of human-agent interaction in open-ended scenarios [2]. The paper synthesizes work on adversarial, backdoor, jailbreak, and hardware-level attacks, alongside research on attack detection, safe training, robust inference, and risk-aware interaction [2]. The safety of embodied AI has drawn increasing attention as large language models and other foundation models are integrated into robotics and autonomous systems [4]. The broader field of AI ethics has long flagged concerns around accountability, transparency, and the potential for physical harm when automated systems influence real-world decisions [5]. Those concerns intensify when AI is given a body. A 2022 survey of AI researchers found that a majority believed there was at least a 10 percent chance that an inability to control AI would lead to an existential catastrophe [3]. In 2023, hundreds of experts signed a statement calling the mitigation of extinction risk from AI a global priority on par with pandemics and nuclear war [3]. The arXiv survey does not engage directly with existential-risk scenarios, but it provides what the authors call a roadmap for building embodied agents that are "safe, robust, and reliable in real-world deployment" [2]. By organizing fragmented lines of research into a single framework, the paper aims to help the field move from capability demonstrations toward systems that can be trusted in transportation, healthcare, industrial robotics, and assistive settings [2].