Let G be a simple algebraic group over an algebraically closed field of characteristic p > 0 and suppose that p is a very good prime for G. In this paper we prove that any maximal Lie subalgebra M of g = Lie(G) with rad(M ) = 0 has the form M = Lie(P ) for some maximal parabolic subgroup P of G. This means that Morozov's theorem on maximal subalgebras is valid under mild assumptions on G. We show that such assumptions are necessary by providing a counterexample to Morozov's theorem for groups of type E8 over fields of characteristic 5. Our proof relies on the main results and methods of the classification theory of finite dimensional simple Lie algebras over fields of prime characteristic. 2010 Mathematics Subject Classification. 17B45. Supported by the Leverhulme Trust (Grant RPG-2013-293).Proof. Replacing L by its p-closure in g if need be we may assume that L is a restricted subalgebra of g. We use induction on the dimension of G. Suppose the statement holds for all standard reductive k-groups of dimension ≤ n (this is obviously true when n = 1). Now suppose dim(G) = n + 1. By Engel's theorem, the Lie algebra [L, L] is nilpotent. If [L, L] = 0 then nil(L) = 0. If [L, L] = 0 then L = L s ⊕ L n where L s is a toral subalgebra of g and L n = nil(L). If L = L s then L ⊆ Lie(T ) for some maximal torus T of G; see [Hum67, Theorem 13.3]. So we are done in this case.Thus we may assume that n := nil(L) = 0. Then z(n) = 0. As [L, L] ⊆ nil(L) by our assumption on L, the adjoint action of L induces a representation of the abelian Lie algebra L/[L, L] on z(n). So there exists a nonzero e ∈ z(n) such that [L, e] ⊆ ke. As e is a nilpotent element of g it admits a cocharacter λ : k × → G optimal in the sense of the Kempf-Rousseau theory. Let P be the parabolic subgroup associated with λ and p = Lie(P ). Then p = i≥0 g(λ, i). By [Pre03, Theorem 2.3], one can choose λ in such a way that e ∈ g(λ, 2) and g e ⊆ p. Furthermore, [g(λ, i), e] = g(λ, i + 2) for all i ≥ 0. Taking i = 0 we find h 0 ∈ g(λ, 0) with [h 0 , e] = e. This implies thatBy our induction assumption, the image of L in the Levi subalgebra g(λ, 0) ∼ = p/nil(p) of g is contained in a Borel subalgebra of g(λ, 0), say b. Since the inverse image of b under the canonical homomorphism p ։ g(λ, 0) is a Borel subalgebra of g, this accomplishes the induction step of our proof.2.4. We denote by O min the minimal nonzero nilpotent orbit in g. It consists of all nonzero e ∈ g with the property that [e, [e, g]] = ke. Corollary 2.3. Let M be a maximal Lie subalgebra of g and denote by N the nilradical of M . Suppose N = 0 and let R be any Lie subalgebra of M whose derived ideal [R, R] consists of nilpotent elements of g. Then the centraliser c g (N ) is an ideal of M and there exists e ∈ c g (N ) ∩ O min such that [R, e] ⊆ ke. Proof. Since N is nilpotent we have that 0 = z(N ) ⊆ c g (N ). Therefore, c g (N ) is a nonzero Lie subalgebra of g. If x ∈ M , c ∈ c g (N ) and n ∈ N then [[x, c], n] = [x, [c, n]] − [c, [x, n]] = 0. So [M, c g (N )] ⊆ c g (N ) which implies that M := M + ...