OpenAI unveiled a collection of adversarial pictures that trick image classifiers even when the viewer changes scale or perspective.
The team generated images that, across a range of zoom levels and viewing angles, are still misidentified by standard neural network classifiers. The work directly counters a recent claim that autonomous‑vehicle cameras, which capture scenes from many viewpoints, would be difficult to deceive.
If self‑driving systems rely on the same visual models, the finding suggests a wider attack surface than previously thought. It gives researchers a concrete test set for hardening perception pipelines, and forces car makers to consider defenses beyond simply adding more cameras.
The demo was posted on July 17, 2017, and includes hundreds of test images. While OpenAI offered no timeline for broader release, the result adds a data point to a growing body of work showing that visual AI remains vulnerable to carefully crafted inputs, even in real‑world‑like conditions.