Witness intimidation is a fundamental threat to the rule of law. It also involves significant strategic complexity and two-sided uncertainty: a criminal cannot know whether his threat will effectively deter a witness from testifying, and a witness cannot know whether the threat will in fact be carried out. We model this interaction and explore the manner in which equilibrium rates of intimidation, testimony, and conviction respond to changes in prosecutorial effectiveness, police-community relations, and witness protection programs. An increase in prosecutorial effectiveness raises the incentives for criminals to threaten witnesses but also makes these threats less credible. Sometimes the rise in threats will be large enough to drive down the rate of conviction, with the paradoxical outcome that better prosecutors may convict fewer criminals. Direct attempts to reduce witness tampering may also prove counterproductive. When the harm faced by a witness itself depends on whether or not the criminal is convicted, communities can be trapped in equilibria with collective silence: no witness testifies because none expects others to testify.