Rival conspecifics often produce stereotyped sequences of signals as agonistic interactions escalate. Successive signals in sequence are thought to convey increasingly pronounced levels of aggressive motivation. Here, we propose and test a model of aggressive escalation in black-throated blue warblers, presenting subjects with two sequential and increasingly elevated levels of threat. From a speaker outside the territorial boundary, we initiated an interaction (low-threat level), and from a second speaker inside the territory, accompanied by a taxidermic mount, we subsequently simulated a territorial intrusion (escalated threat level). Our two main predictions were that signalling behaviours in response to low-threat boundary playback would predict signalling responses to the escalated within-territory threat, and that these latter signalling behaviours would in turn reliably predict attack. We find clear support for both predictions: (i) specific song types (type II songs) produced early in the simulated interaction, in response to boundary playback, predicted later use of low-amplitude 'soft' song, in response to withinterritory playback; and (ii) soft song, in turn, predicted attack of the mount. Unexpectedly, use of the early-stage signal (type II song) itself did not predict attack, despite its apparent role in aggressive escalation. This raises the intriguing question of whether type II song can actually be considered a reliable aggressive signal. Overall, our results provide new empirical insights into how songbirds may use progressive vocal signalling to convey increasing levels of threat.