Penetration testers for AI systems use the Atlas as a checklist. If your model is vulnerable to gradient-based adversarial attacks (Chapter 1), it is not ready for production. The PDF often includes ready-made scripts to test your model against known anomalies.

If you were to download a genuine Atlas of Anomalous AI PDF today, what would you find inside? While versions vary, a canonical atlas contains the following six chapters.

When two or more AIs interact, strange social behaviors emerge:

Unlike adversarial examples, hallucinations occur when the model confidently generates false information. The Atlas maps a typology:

The edges of intelligence reveal the nature of intelligence. When a vision transformer confuses a school bus with an ostrich, it teaches us that the network's internal representation of "bus" is not our own. The Atlas is, therefore, a document of cognitive estrangement.