AI and Robotics: Safety Tests Yield Unexpected Results

Robocurve's RoboHarm experiment tested GPT-6 Astra, Claude Fable 5.1, and MolmoAct2 AI models controlling robots on dangerous physical tasks, revealing that current AI safety mechanisms often fail to prevent risky robot actions. The study highlights gaps in AI safety when models control physical objects versus text-based interactions.

This Wire brief sits within Fusion42's coverage of AI Frontier Models, Robotics and AI Infrastructure.