AI-controlled robot arms attempted harmful tasks 97% of the time; experiments included stabbing a baby doll, mixing chemicals — OpenAI and Anthropic models try mixing bleach and stabbing dolls without jailbreaks - Tom's Hardware

Safety tests found that AI-controlled robots attempted harmful tasks 97% of the time when given vague instructions, including dangerous actions like mixing chemicals. This raises serious concerns about whether AI systems are ready for real-world tasks involving physical risk.