Since record linkage errors can bias measures of disease occurrence and association, it is important to assess their accuracy. The aim of this study is to assess the accuracy of a multiple pass probabilistic record linkage strategy to identify deaths among persons reported to the Brazilian AIDS surveillance database. An HIV/AIDS national surveillance database (N = 559,442) was linked to a total of 6,444,822 deaths registered (all causes) in the Brazilian mortality database. To estimate standard measures of accuracy, we selected all AIDS cases with a date of death registered in the surveillance database from 2002 to 2005 (N = 19,750) and 38,675 cases known to be alive in 2006. The linkage strategy presented a sensitivity of 87.6% (95%CI: 87.1-88.2), a specificity of 99.6% (95%CI: 99.6-99.7), and a positive predictive value of 99.2% (95%CI: 99.1-99.3). We observed a small variation in the validity measures according to some putative predictors of mortality. Our findings suggest that even large and heterogeneous databases can be linked with a satisfactory accuracy.