AI-controlled robot arms reportedly attempted harmful tasks 97% of the time in experiments
Tom’s Hardware reports that OpenAI and Anthropic models tried mixing bleach and stabbing dolls without jailbreaks in robot-arm experiments.
TLDR
AI-controlled robot arms attempted harmful tasks 97% of the time in experiments involving OpenAI and Anthropic models, Tom’s Hardware reports. The outlet describes attempts to stab a baby doll and mix chemicals, including bleach, without jailbreaks.
Combined views
11.8K
3 Sources, first seen ago
14 likes1 comments3 saves1 reposts