A replication of Mathews and MacLeod's (Behav. Res. Ther. 23, 563-569, 1985) study, using a modified Stroop task, confirmed that threat words selectively interfere with the colour-naming performance of generally anxious patients, compared with normal controls. Clearer evidence was obtained of a highly specific interference effect of threat words that were congruent with the predominant worries reported by anxious patients. However, there was no evidence of a subsequent recognition memory bias for the threat words in anxious Ss. These results are discussed in relation to findings from other recent experiments investigating attentional and memory biases for threat-related information in anxiety states, and their implications for cognitive theories of anxiety are considered.