diff --git a/numpy/core/src/multiarray/item_selection.c b/numpy/core/src/multiarray/item_selection.c
index d6e0980c6173..2513bf5900d3 100644
--- a/numpy/core/src/multiarray/item_selection.c
+++ b/numpy/core/src/multiarray/item_selection.c
@@ -87,7 +87,8 @@ PyArray_TakeFrom(PyArrayObject *self0, PyObject *indices0, int axis,
     }
     else {
         int flags = NPY_ARRAY_CARRAY |
-                    NPY_ARRAY_UPDATEIFCOPY;
+                    NPY_ARRAY_UPDATEIFCOPY |
+                    NPY_ARRAY_FORCECAST;
 
         if ((PyArray_NDIM(out) != nd) ||
             !PyArray_CompareLists(PyArray_DIMS(out), shape, nd)) {
@@ -104,8 +105,23 @@ PyArray_TakeFrom(PyArrayObject *self0, PyObject *indices0, int axis,
              */
             flags |= NPY_ARRAY_ENSURECOPY;
         }
+
         dtype = PyArray_DESCR(self);
+        /*
+         * The out array is converted to an array obj of type dtype and after
+         * the take operation is finished is update-if-copy-ed back. An
+         * explicit check that dtype can be safely cast to the type of out
+         * is needed, as the constructor of obj would check the opposite
+         * casting, were it not disabled by the NPY_ARRAY_FORCECAST flag.
+         */
+        if (PyArray_CanCastTypeTo(dtype, PyArray_DESCR(out),
+                                  NPY_SAFE_CASTING) == 0) {
+            PyErr_SetString(PyExc_TypeError,
+                            "output cannot be safely cast to out dtype");
+            goto fail;
+        }
         Py_INCREF(dtype);
+
         obj = (PyArrayObject *)PyArray_FromArray(out, dtype, flags);
         if (obj == NULL) {
             goto fail;
diff --git a/numpy/core/tests/test_item_selection.py b/numpy/core/tests/test_item_selection.py
index d8e9e6fd0faf..048d8da0d43b 100644
--- a/numpy/core/tests/test_item_selection.py
+++ b/numpy/core/tests/test_item_selection.py
@@ -3,6 +3,7 @@
 import numpy as np
 from numpy.testing import *
 import sys, warnings
+import itertools
 
 
 class TestTake(TestCase):
@@ -45,7 +46,6 @@ def test_simple(self):
                             res = ta.take(index_array, mode=mode, axis=1)
                             assert_(res.shape == (2,) + index_array.shape)
 
-
     def test_refcounting(self):
         objects = [object() for i in range(10)]
         for mode in ('raise', 'clip', 'wrap'):
@@ -60,6 +60,17 @@ def test_refcounting(self):
             del a
             assert_(all(sys.getrefcount(o) == 3 for o in objects))
 
+    def test_casting(self):
+        types = ''.join((np.typecodes['AllInteger'], np.typecodes['Float']))
+        modes = ['raise', 'wrap', 'clip']
+        for from_, to_, mode in itertools.product(types, types, modes):
+            a = np.array([0, 1, 2], dtype=np.dtype(from_))
+            b = np.empty((3,), dtype=np.dtype(to_))
+            if np.can_cast(from_, to_):
+                a.take([0, 1, 2], out=b, mode=mode)
+            else:
+                assert_raises(TypeError, a.take, [0, 1, 2], out=b, mode=mode)
+
     def test_unicode_mode(self):
         d = np.arange(10)
         k = b'\xc3\xa4'.decode("UTF8")