The primary dilemma in evolutionarily stable mutualisms is that natural selection for cheating could overwhelm selection for cooperation. Cheating need not entail parasitism; selection favours cheating as a quantitative trait whenever less-cooperative partners are more fit than more-cooperative partners. Mutualisms might be stabilised by mechanisms that direct benefits to more-cooperative individuals, which counter selection for cheating; however, empirical evidence that natural selection favours cheating in mutualisms is sparse. We measured selection on cheating in single-partner pairings of wild legume and rhizobium lineages, which prevented legume choice. Across contrasting environments, selection consistently favoured cheating by rhizobia, but did not favour legumes that provided less benefit to rhizobium partners. This is the first simultaneous measurement of selection on cheating across both host and symbiont lineages from a natural population. We empirically confirm selection for cheating as a source of antagonistic coevolutionary pressure in mutualism and a biological dilemma for models of cooperation.