Mitigating AI Risks: Enhancing Large Language Model Safety in Robot Control

The increasing integration of Large Language Models (LLMs) into robotic systems enables robots to understand and respond to natural language commands. This advancement is particularly valuable in critical sectors such as healthcare, manufacturing, and smart environments, where LLM-driven robots can simplify complex operations and improve accessibility. However, as robots become more reliant on LLMs for decision-making and control, they also become vulnerable to new cybersecurity threats, raising concerns about safety, reliability, and trust in AI-driven automation. This project directly addresses these challenges by developing an innovative AI safety framework to protect LLM-driven robots from adversarial manipulation and unintended harmful behaviours. Existing AI safety measures are not designed to handle sequential command vulnerabilities, adversarial prompts, or indirect manipulation, making LLM-based control systems particularly susceptible to exploitation. Attackers could manipulate prompts, inject harmful instructions, or exploit contextual gaps, leading to unsafe robotic actions. Even without malicious intent, LLMs may misinterpret ambiguous or sequential instructions, posing a risk to users. To mitigate these risks, our project introduces a two-level AI safety guard architecture, integrating edge-level and on-device protection to enhance systemic AI safety in resource-constrained robotic environments. Level 1 Edge-Level Protection: By leveraging edge computing, safety mechanisms operate closer to the source (e.g., within local networks or devices) rather than relying on remote cloud-based servers. This minimises latency, enhances privacy by keeping sensitive data local, and reduces reliance on external infrastructure. Level 2 On-Device Protection: Lightweight AI models embedded directly within robots provide real-time anomaly detection and adversarial resilience. This ensures that even in resource-constrained robotic systems, security mechanisms can function effectively without external dependencies. This dual-layer security model provides robust protection against adversarial attacks, ensuring that even if one layer is compromised or delayed, the other acts as a fail-safe to maintain operational integrity. The key innovations and impact are the following: (1) Context-Aware Safety Mechanisms: Unlike conventional AI security models, our approach ensures that robots understand the intent behind sequential commands, preventing unintended harmful actions. (2) Lightweight AI Security for Edge and Robotics: Many existing cybersecurity solutions are too computationally intensive for real-time robot control. Our model is optimised for efficient deployment in resource-limited robotic environments. (3) Protection Against Emerging AI Threats: By defending against prompt injection, sequential command exploitation, and adversarial manipulation, this project sets a new framework for AI security in robotics.