Strong nonlinear effects combined with diffusive coupling may give rise to unpredictable evolution in spatially extended deterministic dynamical systems even in the presence of a fully negative spectrum of Lyapunov exponents. This regime, denoted as "stable chaos'', has been so far mainly characterized by numerical studies. In this communication we investigate the mechanisms that are at the basis of this form of unpredictable evolution generated by a nonlinear information flow through the boundaries. In order to clarify how linear stability can coexist with nonlinear instability, we construct a suitable stochastic model. In the absence of spatial coupling, the model does not reveal the existence of any self-sustained chaotic phase. Nevertheless, already this simple regime reveals peculiar differences between the behaviour of finite-size and that of infinitesimal perturbations. A mean-field analysis of the truly spatially extended case clarifies that the onset of chaotic behaviour can be traced back to the diffusion process that tends to shift the growth rate of finite perturbations from the quenched to the annealed average. The possible characterization of the transition as the onset of directed percolation could be also briefly discussed as well as the connections with a synchronization transition.