A thin skin of low tensile failure strain, if bonded to the tensile surface of an unnotched impact bend specimen of much tougher material, can change the global failure mode from ductile to brittle. A novel model of this well-known effect is developed and applied to results from impact tests on a tough core of polyamidepolyethylene blend, with a single skin of brittle EVOH. At a fixed crosshead speed, notched specimens of the blend become brittle at a relatively low temperature T bt . Un-notched bilayer specimens continue to show skin fracture up to a considerably higher temperature T fs ; above this temperature they do not fail at all but below T bt they too fail in a brittle manner. Within the temperature range from T fs down to T bt there is a transition from crack arrest, either at the skin/core interface or further into the core where a crack would not normally propagate, to brittle fracture. This brittle fracture temperature is predicted by modelling the process as a three-phase impact event. In the first phase, the striker bends the bilayer quasi-statically. The second phase begins with instantaneous fracture of the skin at its failure strain. The skin ends retract at finite speed, and a craze grows in the adjacent core material to accommodate the local strain singularity. The last phase is a striker-driven impact event similar to that in a notched bend specimen of the core material, except that the crack-tip craze already bears the adiabatic temperature distribution generated while it was driven open by skin retraction. The criterion for craze decohesion, and hence for a crack jump, is the same adiabatic decohesion criterion which accounts for the speed-dependence of impact fracture in notched monolayer specimens. Applied computationally, this model predicts whether a bilayer structure fails in a brittle way or whether cracks initiated in the skin are arrested, either temporarily or permanently, at the skin/core interface.