
AI
Robocurve Finds Frontier Robot Policies Rarely Refuse Unsafe Instructions
In a September 18, 2026 benchmark of 300 robot-arm trials, Robocurve’s RoboHarm report shows Claude Fable 5.1, GPT-6 Astra, and Ai2’s MolmoAct2 generally attempting five fixed harmful instructions without jailbreaks.




