Better safe than sorry: Risk-aware nonlinear Bayesian estimation