▪ Abstract Controversy exists over ecological risks in classical biological control. We reviewed 10 projects with quantitative data on nontarget effects. Ten patterns emerged: (a) Relatives of the pest are most likely to be attacked; (b) host-specificity testing defines physiological host range, but not ecological range; (c) prediction of ecological consequences requires population data; (d) level of impact varied, often in relation to environmental conditions; (e) information on magnitude of nontarget impact is sparse; (f) attack on rare native species can accelerate their decline; (g) nontarget effects can be indirect; (h) agents disperse from agroecosystems; (i) whole assemblages of species can be perturbed; and (j) no evidence on adaptation is available in these cases. The review leads to six recommendations: Avoid using generalists or adventive species; expand host-specificity testing; incorporate more ecological information; consider ecological risk in target selection; prioritize agents; and pursue genetic data on adaptation. We conclude that retrospective analyses suggest clear ways to further increase future safety of biocontrol.