Robust adversarial inputs

In three linesOpenAI created adversarial images that reliably fool neural network classifiers across varied scales and perspectives. This challenges a recent claim that self-driving cars would be hard to trick maliciously due to their multi-angle image capture.

Summary generated by Claude — human-verified