As artificial intelligence moves from computer screens into cars, drones, and robots, a new survey highlights the growing security and ethical risks when these systems interact with the physical world. The review, published in the journal Machine Intelligence Research, examines how vision-language models (VLMs) and vision-language-action models (VLAs) are used in embodied intelligence, where a mistaken description or manipulated command can turn into a physical action.
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, identifies threats across perception, planning, instruction following, and human-robot interaction. These include hallucinations, synthetic forgeries, adversarial attacks, privacy leakage, and unsafe execution. The authors also outline defensive measures, such as contextual checking, forgery detection, and risk-aware reasoning.
“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 no single filter can secure an embodied agent; protection must follow the entire path from sensor input to model reasoning to physical execution.
The survey details how failures can cascade. Biased training data or poor cross-modal alignment can make a model describe objects that are not present. Forged traffic signs or altered labels can misguide perception. Tiny adversarial perturbations or hidden backdoor triggers may bypass safety controls. Persistent sensing can expose identity and location.
To counter these threats, the authors organize defenses into connected layers, including hallucination filtering, cross-modal forgery detection, and privacy-preserving techniques like differential privacy and homomorphic encryption. They also stress the importance of causal explanations and intent alignment so robots can interpret ambiguous instructions and fall back safely.
The implications for developers and regulators are significant. The survey provides a practical checklist for evaluating embodied systems before deployment, addressing technical robustness, regulatory alignment, social equity, and environmental sustainability. This approach could support safer autonomous transport, healthcare assistance, and warehouse automation.
However, the authors warn that strong laboratory results may not transfer to noisy, culturally diverse environments. Progress depends on cross-disciplinary cooperation and testing that measures not only task success but safe behavior under stress.
The review appears in a special issue on the security and ethics of generative AI, with DOI: 10.1007/s11633-025-1626-x. The research was partially supported by the National Natural Science Foundation of China and the Engineering and Physical Sciences Research Council in the UK.
