Higher-order grammars have been extensively studied in 1980's and interests in them have revived recently in the context of higher-order model checking and program verification, where higherorder grammars are used as models of higher-order functional programs. A lot of theoretical questions remain open, however, for unsafe higherorder grammars (grammars without the so-called safety condition). In this paper, we show that any tree languages generated by order-2 unsafe grammars are context-sensitive. This also implies that any unsafe order-3 word languages are context-sensitive. The proof involves novel technique based on typed lambda-calculus, such as type-based grammar transformation. © 2014 Springer-Verlag.