There are rising concerns about the quality of policing in high-violence urban neighborhoods in the United States. Residents of these areas are concerned that police officers are failing to reduce crime, but also that when police officers do engage, their tactics are too severe. This paper examines whether risks to officer safety drive both phenomena. To do so, I exploit variation in unprovoked ambushes on police within and across beats using administrative 911 call data from a large American city. Results show that ambushes lead to an 8 percent decline in arrests, an effect that persists for at least three years after an ambush. In contrast, I find no effect on police severity, as measured by use of force and civilian complaints. This suggests police officers respond to increased risk by de-policing, rather than using more aggressive tactics.