Dr. Shibani Santurkar talks about the limitations of image recognition models and how attackers can manipulate them using adversarial examples, which can segment images in unexpected ways and fool the model into generating false predictions.