diff --git a/Lib/test/output/test_mutants b/Lib/test/output/test_mutants
new file mode 100644
index 00000000000..259c4f56bc5
--- /dev/null
+++ b/Lib/test/output/test_mutants
@@ -0,0 +1 @@
+test_mutants
diff --git a/Lib/test/test_mutants.py b/Lib/test/test_mutants.py
new file mode 100644
index 00000000000..42efb6c88da
--- /dev/null
+++ b/Lib/test/test_mutants.py
@@ -0,0 +1,138 @@
+from test_support import verbose
+import random
+
+# From SF bug #422121:  Insecurities in dict comparison.
+
+# Safety of code doing comparisons has been an historical Python waak spot.
+# The problem is that comparison of structures in written in C *naturally*
+# wants to hold on to things like the size of the container, or "the
+# biggest" containee so far, across a traversal of the container; but
+# code to do containee comparisons can call back into Python and mutate
+# the container in arbitrary ways while the C loop is in midstream.  If the
+# C code isn't extremely paranoid about digging things out of memory on
+# each trip, and artificially boosting refcounts for the duration, anything
+# from infinite loops to OS crashes can result (yes, I use Windows <wink>).
+#
+# The other problem is that code designed to provoke a weakness is usually
+# white-box code, and so catches only the particular vulnerabilities the
+# author knew to protect against.  For example, Python's list.sort() code
+# went thru many iterations as one "new" vulnerability after another was
+# discovered.
+#
+# So the dict comparison test here uses a black-box approach instead,
+# generating dicts of various sizes at random, and performing random
+# mutations on them at random times.  This proved very effective,
+# triggering at least six distinct failure modes the first 20 times I
+# ran it.  Indeed, at the start, the driver never got beyond 6 iterations
+# before the test died.
+
+# The dicts are global to make it easy to mutate tham from within functions.
+dict1 = {}
+dict2 = {}
+
+# The current set of keys in dict1 and dict2.  These are materialized as
+# lists to make it easy to pick a dict key at random.
+dict1keys = []
+dict2keys = []
+
+# Global flag telling maybe_mutate() wether to *consider* mutating.
+mutate = 0
+
+# If global mutate is true, consider mutating a dict.  May or may not
+# mutate a dict even if mutate is true.  If it does decide to mutate a
+# dict, it picks one of {dict1, dict2} at random, and deletes a random
+# entry from it.
+
+def maybe_mutate():
+    if not mutate:
+        return
+    if random.random() < 0.5:
+        return
+    if random.random() < 0.5:
+        target, keys = dict1, dict1keys
+    else:
+        target, keys = dict2, dict2keys
+    if keys:
+        i = random.randrange(len(keys))
+        key = keys[i]
+        del target[key]
+        # CAUTION:  don't use keys.remove(key) here.  Or do <wink>.  The
+        # point is that .remove() would trigger more comparisons, and so
+        # also more calls to this routine.  We're mutating often enough
+        # without that.
+        del keys[i]
+
+# A horrid class that triggers random mutations of dict1 and dict2 when
+# instances are compared.
+
+class Horrid:
+    def __init__(self, i):
+        # Comparison outcomes are determined by the value of i.
+        self.i = i
+
+        # An artificial hashcode is selected at random so that we don't
+        # have any systematic relationship between comparsion outcomes
+        # (based on self.i and other.i) and relative position within the
+        # hawh vector (based on hashcode).
+        self.hashcode = random.randrange(1000000000)
+
+    def __hash__(self):
+        return self.hashcode
+
+    def __cmp__(self, other):
+        maybe_mutate()   # The point of the test.
+        return cmp(self.i, other.i)
+
+    def __repr__(self):
+        return "Horrid(%d)" % self.i
+
+# Fill dict d with numentries (Horrid(i), Horrid(j)) key-value pairs,
+# where i and j are selected at random from the candidates list.
+# Return d.keys() after filling.
+
+def fill_dict(d, candidates, numentries):
+    d.clear()
+    for i in xrange(numentries):
+        d[Horrid(random.choice(candidates))] = \
+            Horrid(random.choice(candidates))
+    return d.keys()
+
+# Test one pair of randomly generated dicts, each with n entries.
+# Note that dict comparison is trivial if they don't have the same number
+# of entires (then the "shorter" dict is instantly considered to be the
+# smaller one, without even looking at the entries).
+
+def test_one(n):
+    global mutate, dict1, dict2, dict1keys, dict2keys
+
+    # Fill the dicts without mutating them.
+    mutate = 0
+    dict1keys = fill_dict(dict1, range(n), n)
+    dict2keys = fill_dict(dict2, range(n), n)
+
+    # Enable mutation, then compare the dicts so long as they have the
+    # same size.
+    mutate = 1
+    if verbose:
+        print "trying w/ lengths", len(dict1), len(dict2),
+    while dict1 and len(dict1) == len(dict2):
+        if verbose:
+            print ".",
+        c = cmp(dict1, dict2)
+    if verbose:
+        print
+
+# Run test_one n times.  At the start (before the bugs were fixed), 20
+# consecutive runs of this test each blew up on or before the sixth time
+# test_one was run.  So n doesn't have to be large to get an interesting
+# test.
+# OTOH, calling with large n is also interesting, to ensure that the fixed
+# code doesn't hold on to refcounts *too* long (in which case memory would
+# leak).
+
+def test(n):
+    for i in xrange(n):
+        test_one(random.randrange(1, 100))
+
+# See last comment block for clues about good values for n.
+test(100)
diff --git a/Misc/NEWS b/Misc/NEWS
index 79066131a4a..8b4021f85d5 100644
--- a/Misc/NEWS
+++ b/Misc/NEWS
@@ -47,6 +47,11 @@ Core
 - Comparing dictionary objects via == and != is faster, and now works even
   if the keys and values don't support comparisons other than ==.
 
+- Comparing dictionaries in ways other than == and != is slower:  there were
+  insecurities in the dict comparison implementation that could cause Python
+  to crash if the element comparison routines for the dict keys and/or
+  values mutated the dicts.  Making the code bulletproof slowed it down.
+
 
 What's New in Python 2.1 (final)?
 =================================
diff --git a/Objects/dictobject.c b/Objects/dictobject.c
index 56cc08f35c4..3f77427aca5 100644
--- a/Objects/dictobject.c
+++ b/Objects/dictobject.c
@@ -981,43 +981,83 @@ PyDict_Items(PyObject *mp)
 
 /* Subroutine which returns the smallest key in a for which b's value
    is different or absent.  The value is returned too, through the
-   pval argument.  No reference counts are incremented. */
+   pval argument.  Both are NULL if no key in a is found for which b's status
+   differs.  The refcounts on (and only on) non-NULL *pval and function return
+   values must be decremented by the caller (characterize() increments them
+   to ensure that mutating comparison and PyDict_GetItem calls can't delete
+   them before the caller is done looking at them). */
 
 static PyObject *
 characterize(dictobject *a, dictobject *b, PyObject **pval)
 {
-	PyObject *diff = NULL;
+	PyObject *akey = NULL; /* smallest key in a s.t. a[akey] != b[akey] */
+	PyObject *aval = NULL; /* a[akey] */
 	int i, cmp;
 
-	*pval = NULL;
 	for (i = 0; i < a->ma_size; i++) {
-		if (a->ma_table[i].me_value != NULL) {
-			PyObject *key = a->ma_table[i].me_key;
-			PyObject *aval, *bval;
-			if (diff != NULL) {
-			    cmp = PyObject_RichCompareBool(diff, key, Py_LT);
-			    if (cmp < 0)
-				return NULL;
-			    if (cmp > 0)
+		PyObject *thiskey, *thisaval, *thisbval;
+		if (a->ma_table[i].me_value == NULL)
+			continue;
+		thiskey = a->ma_table[i].me_key;
+		Py_INCREF(thiskey);  /* keep alive across compares */
+		if (akey != NULL) {
+			cmp = PyObject_RichCompareBool(akey, thiskey, Py_LT);
+			if (cmp < 0) {
+				Py_DECREF(thiskey);
+				goto Fail;
+			}
+			if (cmp > 0 ||
+			    i >= a->ma_size ||
+			    a->ma_table[i].me_value == NULL)
+			{
+				/* Not the *smallest* a key; or maybe it is
+				 * but the compare shrunk the dict so we can't
+				 * find its associated value anymore; or
+				 * maybe it is but the compare deleted the
+				 * a[thiskey] entry.
+				 */
+				Py_DECREF(thiskey);
 				continue;
 			}
-			aval = a->ma_table[i].me_value;
-			bval = PyDict_GetItem((PyObject *)b, key);
-			if (bval == NULL)
-				cmp = 0;
-			else {
-			    cmp = PyObject_RichCompareBool(aval, bval, Py_EQ);
-			    if (cmp < 0)
-				return NULL;
-			}
-			if (cmp == 0)
-			{
-				diff = key;
-				*pval = aval;
+		}
+
+		/* Compare a[thiskey] to b[thiskey]; cmp <- true iff equal. */
+		thisaval = a->ma_table[i].me_value;
+		assert(thisaval);
+		Py_INCREF(thisaval);   /* keep alive */
+		thisbval = PyDict_GetItem((PyObject *)b, thiskey);
+		if (thisbval == NULL)
+			cmp = 0;
+		else {
+			/* both dicts have thiskey:  same values? */
+			cmp = PyObject_RichCompareBool(
+						thisaval, thisbval, Py_EQ);
+			if (cmp < 0) {
+		    		Py_DECREF(thiskey);
+		    		Py_DECREF(thisaval);
+		    		goto Fail;
 			}
 		}
+		if (cmp == 0) {
+			/* New winner. */
+			Py_XDECREF(akey);
+			Py_XDECREF(aval);
+			akey = thiskey;
+			aval = thisaval;
+		}
+		else {
+			Py_DECREF(thiskey);
+			Py_DECREF(thisaval);
+		}
 	}
-	return diff;
+	*pval = aval;
+	return akey;
+
+Fail:
+	Py_XDECREF(akey);
+	Py_XDECREF(aval);
+	*pval = NULL;
+	return NULL;
 }
 
 static int
@@ -1031,19 +1071,40 @@ dict_compare(dictobject *a, dictobject *b)
 		return -1;	/* a is shorter */
 	else if (a->ma_used > b->ma_used)
 		return 1;	/* b is shorter */
+
 	/* Same length -- check all keys */
+	bdiff = bval = NULL;
 	adiff = characterize(a, b, &aval);
-	if (adiff == NULL && PyErr_Occurred())
-		return -1;
-	if (adiff == NULL)
-		return 0;	/* a is a subset with the same length */
+	if (adiff == NULL) {
+		assert(!aval);
+		/* Either an error, or a is a subst with the same length so
+		 * must be equal.
+		 */
+		res = PyErr_Occurred() ? -1 : 0;
+		goto Finished;
+	}
 	bdiff = characterize(b, a, &bval);
-	if (bdiff == NULL && PyErr_Occurred())
-		return -1;
-	/* bdiff == NULL would be impossible now */
-	res = PyObject_Compare(adiff, bdiff);
-	if (res == 0)
+	if (bdiff == NULL && PyErr_Occurred()) {
+		assert(!bval);
+		res = -1;
+		goto Finished;
+	}
+	res = 0;
+	if (bdiff) {
+		/* bdiff == NULL "should be" impossible now, but perhaps
+		 * the last comparison done by the characterize() on a had
+		 * the side effect of making the dicts equal!
+		 */
+		res = PyObject_Compare(adiff, bdiff);
+	}
+	if (res == 0 && bval != NULL)
 		res = PyObject_Compare(aval, bval);
+
+Finished:
+	Py_XDECREF(adiff);
+	Py_XDECREF(bdiff);
+	Py_XDECREF(aval);
+	Py_XDECREF(bval);
 	return res;
 }