A set of vertices S is a resolving set of a graph G; if for every pair of vertices x and y in G, there exists a vertex s in S such that x and y differ in distance to s. A smallest resolving set of G is called a metric basis. The metric dimension dim(G) is the cardinality of a metric basis of G. The notion of a metric basis is applied to the problem of placing sensors in a network, where the problem of sensor faults can arise. The fault-tolerant metric dimension ftdim(G) is the cardinality of a smallest resolving set S such that S\{s} remains a resolving set of G for every s in S. A natural question is how much more sensors need to be used to achieve a fault-tolerant metric basis. It is known in literature that there exists an upper bound on ftdim(G) which is exponential in terms of dim(G); i.e. ftdim(G) <= dim(G)(1+2^(5dim(G)-1)). In this paper, we construct graphs G with ftdim(G) = dim(G)+2^(dim(G)-1) for any value of dim(G), so the exponential upper bound is necessary. We also extend these results to the k-metric dimension which is a generalization of the fault-tolerant metric dimension. First, we establish a similar exponential upper bound on dim(k+1)(G) in terms of dim(k)(G); and then we show that there exists a graph for which dim(k+1)(G) is indeed exponential. For a possible further work, we leave the gap between the bounds to be reduced.