The security of AI environments cannot be based on the measures of the perimeter alone; it must be verified on a regular basis.