Researchers at OpenAI have developed a method to test how well neural network classifiers can defend against unexpected attacks. They've created a new metric called UAR (Unforeseen Attack Robustness) that measures the robustness of a single model against an unanticipated attack. This highlights the need for more diverse testing, as current methods may not be enough to prepare models for real-world threats. --- Why it matters: This matters because AI systems are increasingly being deployed in critical applications where security is paramount. Developing robust methods to test their defenses against unforeseen attacks is crucial to preventing potential failures or even malicious exploitation. Source: https://openai.com/index/testing-robustness