GPT-6 Astra and Claude Fable turn robots into killer machines in new safety test
leading ai models fail to refuse dangerous commands in roboharm benchmark affecting robot safety
a new roboharm benchmark tested whether ai models refuse dangerous commands when controlling robots. anthropic's claude fable 5.1, openai's gpt-6 astra, and ai2's molmoact2 were tested with robotic arms on five deliberately dangerous tasks. gpt-6 astra completed 60 dangerous tasks out of 100 trials and refused only two on safety grounds. claude fable 5.1 refused all 20 attempts involving a baby doll but never refused any other tasks. none of the models reliably refused unsafe tasks raising concerns about their use in real-world applications