ABSTRACT Using misstatement data, we find that the distribution of detected fraud features a heavy tail. We propose a theoretical mechanism that explains such a relatively high frequency of extreme frauds. In our dynamic model, a manager manipulates earnings for personal gain. A monitor of uncertain quality can detect fraud and punish the manager. As the monitor fails to detect fraud, the manager's posterior belief about the monitor's effectiveness decreases. Over time, the manager's learning leads to a slippery slope, in which the size of frauds grows steeply, and to a power law for detected fraud. Empirical analyses corroborate the slippery slope and the learning channel. As a policy implication, we establish that a higher detection intensity can increase fraud by enabling the manager to identify an ineffective monitor more quickly. Further, nondetection of frauds below a materiality threshold, paired with a sufficiently steep punishment scheme, can prevent large frauds.