Artefacts that differentiate spoofed from bona-fide utterances can reside inspectral or temporal domains. Their reliable detection usually depends uponcomputationally demanding ensemble systems where each subsystem is tuned tosome specific artefacts. We seek to develop an efficient, single system thatcan detect a broad range of different spoofing attacks without score-levelensembles. We propose a novel heterogeneous stacking graph attention layerwhich models artefacts spanning heterogeneous temporal and spectral domainswith a heterogeneous attention mechanism and a stack node. With a new max graphoperation that involves a competitive mechanism and an extended readout scheme,our approach, named AASIST, outperforms the current state-of-the-art by 20%relative. Even a lightweight variant, AASIST-L, with only 85K parameters,outperforms all competing systems.