Humans routinely incur costs and take risks to help others. Current theories across the behavioral sciences explain prosocial behavior by assuming that people are rational but altruistic. However, extant empirical tests of these theories have focused on riskless prosocial behaviors and implicitly assumed their findings will also apply to risky prosocial behavior. We challenge this assumption. Across seven field and laboratory studies, we show that risk transforms prosocial behavior. In a field setting, we examine millions of risky prosocial behaviors by volunteer crisis counselors who try to help others deal with depression, abuse, and suicide. By leveraging plausibly exogenous increases in positive affect (driven by good news, happy calendar days, and sunshine), we show that unlike riskless prosocial behavior, risky prosocial behavior actually decreases when people experience an exogenous increase in positive affect. These findings contradict current theories and lay beliefs. In pre-registered experiments, we show that loss-frame (vs. gain-frame) nudges boost risky prosocial behavior by 38%-40%, but do not influence riskless prosocial behavior. These findings contradict current theories and identify a novel prosocial-behavior nudge that is nearly ten times stronger than leading prosocial-behavior nudges. To explain these findings, we argue that deciding to engage in risky prosocial behavior creates a warm glow gamble – prosocial successes feel good, but prosocial failures feel bad – which people evaluate in line with Prospect Theory. Our Warm Glow Gamble framework parsimoniously explains our findings. Because many real-world prosocial behaviors are risky, the largest barrier to prosocial behavior may not be selfishness, but risk aversion.