With companies enrolling AI to help look over the shoulder of their employees, I wonder how hard it would be to do some prompt injection just changing what is displayed in the surveiled screen for it to see. Potential for a new vulnerability vector?