This work proposes strategies to handle three types of constraints in the context of blackbox optimization: binary constraints that simply indicate if they are satisfied or not; unrelaxable constraints that are required to be satisfied to trust the output of the blackbox; hidden constraints that are not explicitly known by the user but are triggered unexpectedly. Using tools from classification theory, we build surrogate models of those constraints to guide the MADS algorithm. Numerical results are conducted on three engineer problems.