Machines don't have to become superintelligent to cause catastrophic harm. As artificial intelligence spreads through military decision-making, they just have to be wrong—and trusted by the humans in charge.