Researchers at OpenAI have created images that consistently fool neural network classifiers when viewed from different scales and perspectives. This finding challenges a recent claim that self-driving cars are resistant to malicious inputs due to their ability to capture images from multiple angles. The robust adversarial inputs were designed to evade detection by state-of-the-art image classification models, highlighting the ongoing challenge of developing secure AI systems. --- Why it matters: This matters because it shows that current neural network classifiers can still be fooled with carefully crafted inputs, which has implications for applications like self-driving cars and other safety-critical systems. Source: https://openai.com/index/robust-adversarial-inputs