Stereotype threat (ST) occurs when the awareness of a negative stereotype about a social group in a particular domain produces suboptimal performance by members of that group. Although ST has been repeatedly demonstrated, far less is known about how its effects are realized. Using mathematical problem solving as a test bed, the authors demonstrate in 5 experiments that ST harms math problems that rely heavily on working memory resources--especially phonological aspects of this system. Moreover, by capitalizing on an understanding of the cognitive mechanisms by which ST exerts its impact, the authors show (a) how ST can be alleviated (e.g., by heavily practicing once-susceptible math problems such that they are retrieved directly from long-term memory rather than computed via a working-memory-intensive algorithm) and (b) when it will spill over onto subsequent tasks unrelated to the stereotype in question but dependent on the same cognitive resources that stereotype threat also uses. The current work extends the knowledge of the causal mechanisms of stereotype threat and demonstrates how its effects can be attenuated and propagated.