Despite their strong performance, deep networks remain fragile to adversarial perturbations, often because they rely on imperceptible, non-robust features. Yet there is no principled method to localize the boundary between non-robust and robust feature regimes. We introduce a curvature-based framework that identifies this transition via a continuous pixel-budget attack. By perturbing inputs in saliency-ranked order and analyzing the second derivative of adversarial-accuracy curves, we define a formal “knee” that marks the exhaustion of non-robust features. Across five ImageNet models, the “knee” consistently aligns with a shift in transferability dynamics: models with earlier knees produce more universal attacks, whereas models with later knees better preserve accuracy. Our method establishes a reproducible boundary between brittle and stable representations, offering a new axis for evaluating and comparing adversarial robustness across architectures.