We consider the safety problem for Administrative Role-Based Access Control (ARBAC) policies, i.e. detecting whether sequences of administrative actions can result in policies by which a user can acquire permissions that may compromise some security goals. In particular, we are interested in sequences of safety problems generated by modifications (namely, adding/deleting an element to/from the set of possible actions) to an ARBAC policy accommodating the evolving needs of an organization. or resulting from fixing some safety issues. Since problems in such sequences share almost all administrative actions, we propose an incremental technique that avoids the re-computation of the solution to the current problem by re-using much of the work done on the previous problem in a sequence. An experimental evaluation shows the better performances of an implementation of our technique with respect to the only available approach to solve safety problems for evolving ARBAC policies proposed by Gofman, Luo, and Yang.