As artificial intelligence moves from digital interfaces into cars, drones, and service robots, a new survey highlights the critical security and ethical challenges that emerge when these systems interact with the physical world. The review, published in Machine Intelligence Research, examines how vision-language models (VLMs) and vision-language-action models (VLAs) are used in embodied intelligence, where perception, reasoning, and action combine to guide autonomous systems. The findings underscore that even minor errors or malicious inputs can lead to dangerous outcomes, such as collisions or equipment damage.
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 key vulnerabilities across four core functions: perception, planning, instruction following, and human-robot interaction. These vulnerabilities include hallucinations, where the model describes objects that are not present; synthetic forgeries, such as fake traffic signs or altered labels; adversarial attacks that subtly manipulate inputs to mislead the system; privacy leakage from continuous sensing; and unsafe execution of commands.
According to the authors, the interconnected nature of these systems means that a failure in one layer can cascade into others. For example, biased training data can lead to poor visual-language alignment, causing the model to misinterpret its surroundings. Maliciously crafted inputs, such as forged traffic signs or cloned voices, can then exploit these weaknesses to redirect decisions. The review organizes countermeasures into equally connected layers, including hallucination filtering, cross-modal forgery detection, defenses against adversarial perturbations and backdoors, privacy-preserving techniques like differential privacy and homomorphic encryption, and safeguards for physical control.
The central insight of the review is that no single defense mechanism can secure an embodied agent. Instead, protection must span the entire path from sensor input to model reasoning, system architecture, and physical execution. The authors advocate for combining multiple defense strategies, transparent risk metrics, continuous monitoring, and human oversight for critical decisions. A trustworthy robot must also be able to explain its actions, recognize uncertainty, and fall back safely when conditions are ambiguous.
The review provides a practical checklist for developers and regulators evaluating embodied systems before large-scale deployment. It emphasizes the need for designs that address technical robustness, regulatory alignment, social equity, and environmental sustainability. The authors warn that strong laboratory results may not translate to real-world settings, which are noisy, culturally diverse, and resource-constrained. Therefore, progress depends on cross-disciplinary cooperation and testing that measures not only task success but also safe behavior under stress.
This research is particularly timely as embodied AI systems are increasingly deployed in autonomous transport, healthcare, warehouse automation, and collaborative robotics. By mapping the risks and defenses, the survey offers a roadmap for developing AI that is not only capable but also dependable and safe in physical environments, ensuring that responsibility can be traced when failures occur.
