# Researchers Create Images That Fool AI From Multiple Angles

OpenAI announced they've developed adversarial images capable of reliably tricking neural network classifiers even when viewed from different scales and perspectives.

This breakthrough challenges recent assertions about the security of self-driving car vision systems. Last week, researchers claimed autonomous vehicles would be difficult to deceive maliciously because they capture images from multiple angles, distances, and viewpoints simultaneously.

The implication was that adversarial attacks—carefully crafted images designed to fool AI systems—wouldn't work in real-world driving scenarios where cameras observe objects from various positions.

OpenAI's new research demonstrates this assumption may be overly optimistic. Their robust adversarial inputs maintain their deceptive properties across the varied viewing conditions that self-driving cars encounter.

This matters significantly for autonomous vehicle safety. If attackers can create physical objects or modifications that consistently fool AI vision systems regardless of viewing angle,