We model censorship as a dynamic game between an agent and an evaluator. Two types of public news, good and bad news, are informative about the agent’s ability. However, the agent can hide bad news from the evaluator, at some cost, and will do so if and only if this secures her a significant increase in tenure. Thus, the evaluator faces a bandit problem with endogenous news processes. When bad news is conclusive, the agent always censors when the public belief is sufficiently high, but below a threshold, she entirely or partially stops censoring. The possibility of censorship hurts the evaluator and the good agent, and it may also hurt the bad agent. However, when bad news is inconclusive, we show that the good agent censors bad news more aggressively than the bad agent does. This improves the quality of public information and may benefit all players.