AI‑controlled robots can be fooled by distorted images into unsafe actions

Robocurve conducted two studies that showed AI models controlling real robotic manipulators rarely refuse dangerous tasks, unlike their behavior in online chat. The models were more likely to attempt unsafe actions when the robot's camera feed was blurred or deliberately altered.

In September, Robocurve released the RoboHarm benchmark results, testing Claude Fable 5.1 (Anthropic), GPT‑6 Astra (OpenAI) and MolmoAct2 (Ai2) on two‑handed I2RT YAM manipulators.