A prominent hypothesis proposes that pathogen virulence evolves in large part due to a trade-off between infectiousness and damage to hosts. Other explanations emphasize how virulence evolves in response to competition among pathogens within hosts. Given the proliferation of theoretical possibilities, what best predicts how virulence evolves in real biological systems? Here, I show that virulence evolution in experimental populations of bacteria and self-transmissible plasmids is best explained by within-host competition. Plasmids evolved to severely reduce the fitness of their hosts even in the absence of uninfected cells. This result is inconsistent with the trade-off hypothesis, which predicts that under these conditions vertically transmitted pathogens would evolve to be less virulent. Plasmid virulence was strongly correlated with the ability to superinfect cells containing competing plasmid genotypes, suggesting a key role for within-host competition. When virulent genotypes became common, hosts evolved resistance to plasmid infection. These results show that the trade-off hypothesis can incorrectly predict virulence evolution when within-host interactions are neglected. They also show that symbioses between bacteria and plasmids can evolve to be surprisingly antagonistic.