· Policy & regulation· 日本語で読む
Robocurve Publishes Robot Evaluation Against Dangerous Instructions
Independent evaluation organization Robocurve published results from the RoboHarm benchmark on September 18, 2026.
What happened
Robocurve announced results from RoboHarm, a benchmark that uses a dual-arm robot to examine responses to dangerous instructions. According to reports, OpenAI's GPT-6 Astra, Anthropic's Claude Fable 5.1 and Ai2's MolmoAct2 were tested.
Sources
Translated from the Japanese original: /news/robocurve-souwan-robotto-muke-anzenkijun-roboharm-wo-kouhyou-f4c767. The Japanese version is authoritative if they differ.