Hacking News
Robocurve Tests AI Models with Real Robot Arms, Finds Safety Issues
Robocurve, an independent AI and robotics evaluation firm, tested five frontier AI models on real robot arms and found that they were often compliant with unsafe tasks. The firm released its findings, titled 'Perceptual Attacks on Embodied AI,' in early November 2026. The study revealed that when the robot's camera feed was blurred or doctored, the models were more likely to attempt unsafe tasks, highlighting critical safety concerns in AI-controlled robotics.