Gaps in Robot Intelligence: Study Reveals Robots Respond to Dangerous Commands in 97% of Tests

A study by the American company RoboCare reveals that AI models controlling robots responded to dangerous commands in 97% of tests when presented with modified images, compared to 27% when using original images, raising security concerns as robot use expands.

A recent study conducted by the American company RoboCare, specializing in robot safety assessment, has uncovered security vulnerabilities in AI models that control robots. The study showed that these models attempted to perform unsafe tasks in 97% of tests when presented with modified images, compared to 27% when using original images. These vulnerabilities can be exploited by distorting images captured by robot cameras or displaying written instructions in front of them, without needing to hack their software systems. These findings come at a time when technology companies are expanding the development of humanoid robots and automated systems relying on artificial intelligence, raising questions about the ability of current protection systems to prevent robots from performing actions that could endanger humans.