OpenAI has produced images that fool neural network classifiers no matter the scale or perspective from which they are captured.
The research involves adversarial inputs — images deliberately crafted to make AI vision systems misclassify what they see. What distinguishes this work is robustness across viewing conditions: the images remain effective when captured at different distances, angles, and orientations. That matters because a popular rebuttal to adversarial attacks on autonomous vehicles has been that their camera arrays see the world from multiple simultaneous viewpoints, neutralizing any single manipulated image. That argument appeared explicitly just days before this research was published. OpenAI's images appear to defeat it directly.
The stakes are not abstract. Self-driving cars depend on computer vision to identify stop signs, pedestrians, lane markings, and obstacles. If an adversary can place an image — on a billboard, a road sign, or painted pavement — that reliably fools the classifier regardless of which camera picks it up, the vehicle's redundant sensor array offers no protection. The multi-angle defense was one of the more credible arguments for why adversarial manipulation of autonomous vehicles was a manageable, bounded risk. That argument now looks considerably weaker.
Adversarial examples have been a documented problem in machine learning for over a decade. What recurs reliably is the industry's habit of declaring them contained — until the next paper drops.