The use of vision as a primary sensor for autonomous vehicles like Tesla's autopilot may be vulnerable to adversarial machine learning attacks, where attacks on the imitation model could be transferred to the original model.