As artificial intelligence (AI) transitions from digital interfaces to physical systems like autonomous vehicles, drones, and service robots, the stakes escalate: a misperception or manipulated command can result in real-world harm. A new survey published in Machine Intelligence Research provides a detailed map of the security and ethical risks inherent in embodied AI, where vision-language models (VLMs) and vision-language-action models (VLAs) guide physical actions. The review, 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, synthesizes current threats and defenses, offering a roadmap for developing dependable embodied intelligence.
The integration of VLMs and VLAs enables robots to understand natural language instructions, interpret visual scenes, and execute tasks flexibly. However, this creates a chain of dependency where flawed data, weak alignment, or malicious inputs can cascade into unsafe actions. In a chatbot, such errors might produce misinformation; in an autonomous vehicle, they could lead to collisions. The survey identifies major threats including hallucinations—where models describe objects that do not exist—synthetic forgeries like fake traffic signs or cloned voices, adversarial attacks that exploit tiny perturbations, backdoor triggers, and jailbreak prompts. Privacy leakage through persistent sensing is another concern, as robots may inadvertently collect sensitive data about individuals' identities, locations, and behaviors.
To counter these risks, the authors organize defensive strategies into interconnected layers, aligning with the perception-planning-action pipeline. These include hallucination filtering, cross-modal forgery detection, watermarking, and defenses against adversarial perturbations. Privacy-preserving techniques such as differential privacy, secure multi-party computation, and homomorphic encryption are also discussed. Additionally, the review emphasizes the importance of interpretable reasoning, intent alignment, and risk assessment to enable robots to handle ambiguous instructions and anticipate hazards. The central insight is that no single safeguard suffices; protection must span the entire path from sensor input to physical execution.
The authors stress that the goal is not merely improving accuracy but ensuring safety under imperfect conditions. They advocate for combining defenses rather than deploying isolated patches, with transparent risk metrics and human oversight for critical decisions. A trustworthy robot must be able to explain its actions, recognize uncertainty, and revert to safe fallback behaviors. The review also calls for addressing technical robustness alongside regulatory alignment, social equity, and environmental sustainability, noting that strong lab results may not translate to noisy, diverse real-world environments.
For developers and regulators, this survey serves as a practical checklist for evaluating embodied systems before deployment. Future platforms could integrate interpretable reasoning, attack detection, and privacy-preserving computation under open evaluation protocols. The authors hope this roadmap will support safer autonomous transport, healthcare assistance, warehouse automation, and collaborative robotics, while making responsibility traceable when failures occur. As AI increasingly operates in the physical world, ensuring its safety and ethical integrity becomes paramount, and this review provides a critical foundation for that endeavor.
