AI Safety in the Physical World: New Survey Maps Risks and Defenses for Embodied Intelligence

As artificial intelligence extends beyond digital interfaces into cars, drones, and robots, ensuring safety becomes a critical challenge. A comprehensive review, published in the journal Machine Intelligence Research, examines the security and ethical risks when vision-language models (VLMs) and vision-language-action models (VLAs) guide these embodied systems. The study highlights that a mistaken description or manipulated command can lead to physical actions with real-world consequences.

The survey, conducted by researchers from the Institute of Automation, Chinese Academy of Sciences, University College London, Minzu University of China, and the China Academy of Electronics and Information Technology, maps how failures in perception, planning, instruction following, and human-robot interaction can cascade. It covers threats such as hallucinations, synthetic forgeries, adversarial attacks, privacy leakage, and unsafe execution. The authors argue that existing safeguards are often fragmented and insufficient for real-time, uncertain environments.

“The central challenge is not simply making models more accurate, but ensuring that a system remains safe when its sensors, language inputs, and operating conditions are imperfect,” the authors note. They emphasize that defenses must be combined across the entire pipeline—from sensor input to model reasoning, system architecture, and physical execution. This includes hallucination filtering, cross-modal forgery detection, adversarial defenses, differential privacy, and risk-aware reasoning.

The review also addresses broader implications for accountability, fairness, privacy, and environmental sustainability. For developers and regulators, it provides a practical checklist for evaluating embodied systems before deployment. Future platforms could integrate interpretable reasoning, attack detection, and dynamic safety controls under open evaluation protocols.

The authors call for designs that address technical robustness, regulatory alignment, social equity, and environmental sustainability together. Such an approach could support safer autonomous transport, healthcare assistance, and industrial automation, while making responsibility easier to trace when failures occur. However, they caution that strong laboratory results may not transfer well to noisy, culturally diverse, and resource-constrained environments, stressing the need for cross-disciplinary cooperation and stress testing.

The full study, with DOI 10.1007/s11633-025-1626-x, appears in a special issue on the security and ethics of generative AI. The research was partially supported by the National Natural Science Foundation of China and the Engineering and Physical Sciences Research Council, UK.