Robocurve, a nonprofit, introduced RoboHarm, a benchmark testing how AI models controlling robot arms respond to dangerous commands across five scenes involving a doll and a knife, heating appliances, electrical items, batteries, and cleaning products' packaging materials. GPT-6 Astra tried to execute the dangerous command 97 times out of 100 attempts and succeeded in 60 attempts, estimated by researchers to be about 62% of the attempts. Claude Fable 5.1 declined far more often, attempting only 80% and completing 34% of attempts, mostly in the doll and knife scene. Co-author Jay Chooi highlighted the results on X. According to Robocurve's GitHub method, all models were tested by a single instruction for each scene through twenty independent resetting trials, with reviewers evaluating behavior using camera recording and saved transcripts. Some instructions required indirect referencing, obliging scene analysis. Researchers emphasized instructed behavior, not the machine's ability to generate dangerous intentions. The rubric distinguishes refusal from failed execution, and RoboHarm's scope is deliberately limited to five fixed scenes and single instructions. Elon Musk reposted the findings with a two-word reaction: "Sounds bad."