AI-powered robots are 
becoming essential in areas like healthcare, transportation, and industry. They can perform complex tasks and make decisions using large language models (LLMs). While these advances offer convenience, they also bring serious risks. Recent research reveals that AI robots can be tricked into doing harmful things, raising concerns about public safety and security.
Robots that rely on large LLMs can be tricked with commands if the instructions are cleverly phrased. Hackers might exploit this flaw to make robots behave in dangerous ways. For example, a self-driving car could be tricked into ignoring stop signs or driving into restricted areas. A robot dog might be made to spy on people, and robotic arms could be manipulated to throw objects. This occurs when language prompts deceive the system’s built-in safety checks.
Some Researchers Behind This Discovery
- George Pappas: Head of a research lab from the University of Pennsylvania who studies how LLMs interact with the physical actions of robots.
- Yi Zeng: A PhD student from the University of Virginia, who emphasizes the need for stronger safeguards to prevent robot misuse.
- Pulkit Agrawal: An MIT professor, focused on how robots’ physical movements, like those of robotic arms, can be influenced by AI prompts.
Large language models (LLMs) can produce harmful outputs like offensive language or dangerous instructions because they predict text based on patterns. Human testers typically fine-tune them to avoid this, but clever prompts can still bypass these safeguards. Researchers developed “jailbreak” techniques where robots are tricked with imaginative prompts, like pretending to be characters from movies or video games, leading them to perform unethical tasks. Since LLMs are being used in critical systems like self-driving cars and medical devices, these vulnerabilities pose serious safety and ethical concerns.
AI models that process images, speech, and sensor data are vulnerable to new hacking methods. These inputs can be misused to make robots act unpredictably. According to Alex Robey, a Carnegie Mellon researcher, using video, images, and speech to interact with AI increases possible points of attack, making the system more exposed to manipulation.