We study whether regulators should reveal the models they use to stress test banks. In our setting, revealing leads to gaming, but not revealing can induce banks to underinvest in socially desirable assets for fear of failing the test. We show that although the regulator can solve this underinvestment problem by making the test easier, some disclosure may still be optimal, which under some conditions takes the simple form of a cutoff rule. We characterize the optimal disclosure policy combined with the test difficulty, provide comparative statics, and relate our results to recent policies. We also offer applications beyond stress tests.