How does dictatorship influence scientific production? While much historical work documents how autocratic regimes crack down on institutions of higher learning, we know little about the general patterns linking autocracy to universities and their output (research). We propose what we call a ``security-tradeoff" argument to explain variation in scientific production across regime types. This argues that security concerns weigh heavy when dictators choose to invest in or crack down on academic activity. Dictators must trade potential economic and technological gains from science off against potential security consequences for the regime and that scientists adapt. We proceed to test implications of this argument in different datasets combinining information on political regimes, 13 000 universities and more than 20 million scientific publications from the Web of Science, to study patterns of scientific production 1950-2020. Using synthetic diff-in-diff estimators, we show country-year evidence that autocracies significantly hamper aggregate scientific output, and disproportionally harm social sciences and the humanities. We also study scientific production at subnational and local levels, where we estimate university-fixed effects models showing that academic fields are differentially harmed by autocratization, and grid-cell models showing that science is adversely affected in local areas where regime-opposition is strong.