This research examines critical security vulnerabilities in robotic systems that rely on large language and vision-language models, revealing how seemingly minor input variations can lead to significant operational failures.

• Input Sensitivity: LLM/VLM-controlled robots show inconsistent performance when faced with slight variations in instructions or visual inputs
• Misalignment Issues: These sensitivities trigger misalignments between model interpretations and intended commands
• Safety Implications: Execution failures can have severe real-world consequences, especially in sensitive or high-risk environments
• Security Imperative: Findings emphasize the urgent need for enhanced robustness testing and safety protocols in AI-powered robotics

This research is particularly relevant for security professionals as it highlights significant risks in next-generation robotic systems that must be addressed before widespread deployment in critical applications.

On the Vulnerability of LLM/VLM-Controlled Robotics