diff --git a/Lib/multiprocessing/reduction.py b/Lib/multiprocessing/reduction.py
index fcccd3eef86cc7e..41de207ebb31d47 100644
--- a/Lib/multiprocessing/reduction.py
+++ b/Lib/multiprocessing/reduction.py
@@ -105,6 +105,20 @@ class DupHandle(object):
         '''Picklable wrapper for a handle.'''
         def __init__(self, handle, access, pid=None):
             if pid is None:
+                popen = context.get_spawning_popen()
+                if popen is not None:
+                    # gh-154208: The child process is already running, so
+                    # duplicate the handle straight into it.  Stealing the
+                    # handle (below) would leak the duplicate held by this
+                    # process if the child died before it could steal it.
+                    # This mirrors the fix made for the pipe used to send
+                    # the child its parameters -- see bpo-33929.
+                    self._handle = _winapi.DuplicateHandle(
+                        _winapi.GetCurrentProcess(),
+                        handle, popen.sentinel, access, False, 0)
+                    self._access = access
+                    self._pid = popen.pid
+                    return
                 # We just duplicate the handle in the current process and
                 # let the receiving process steal the handle.
                 pid = os.getpid()
diff --git a/Lib/test/_test_multiprocessing.py b/Lib/test/_test_multiprocessing.py
index 8f665b3a98a0371..821b49d28537d1e 100644
--- a/Lib/test/_test_multiprocessing.py
+++ b/Lib/test/_test_multiprocessing.py
@@ -5733,6 +5733,71 @@ def test_neg_timeout(self):
         a.close()
         b.close()
 
+#
+# Issue 154208: handles duplicated for a child process must not leak in the
+# parent if the child dies before it can make use of them
+#
+
+@unittest.skipUnless(WIN32, "skipped on non-Windows platforms")
+class TestWindowsHandleLeak(unittest.TestCase):
+
+    @staticmethod
+    def handle_count():
+        import ctypes
+        from ctypes import wintypes
+
+        kernel32 = ctypes.WinDLL('kernel32', use_last_error=True)
+        kernel32.GetCurrentProcess.argtypes = ()
+        kernel32.GetCurrentProcess.restype = wintypes.HANDLE
+        kernel32.GetProcessHandleCount.argtypes = (wintypes.HANDLE,
+                                                  wintypes.LPDWORD)
+        kernel32.GetProcessHandleCount.restype = wintypes.BOOL
+
+        # Pseudo-handle: it must not be closed
+        hproc = kernel32.GetCurrentProcess()
+        count = wintypes.DWORD()
+        if not kernel32.GetProcessHandleCount(hproc, ctypes.byref(count)):
+            raise ctypes.WinError(ctypes.get_last_error())
+        return count.value
+
+    @staticmethod
+    def _sleep(conn):
+        time.sleep(support.LONG_TIMEOUT)
+
+    def spawn_and_terminate(self):
+        # Kill the child before it gets a chance to unpickle its arguments:
+        # the handles duplicated for it are then never claimed.
+        reader, writer = multiprocessing.Pipe(duplex=False)
+        try:
+            proc = multiprocessing.Process(target=self._sleep, args=(reader,))
+            proc.start()
+            try:
+                proc.terminate()
+                proc.join()
+            finally:
+                proc.close()
+        finally:
+            reader.close()
+            writer.close()
+
+    def test_no_handle_leak_if_child_is_terminated(self):
+        # Warm up: the first iterations import modules and populate caches,
+        # which opens handles that are unrelated to this test.
+        for _ in range(3):
+            self.spawn_and_terminate()
+        gc.collect()
+
+        before = self.handle_count()
+        for _ in range(10):
+            self.spawn_and_terminate()
+        gc.collect()
+        leaked = self.handle_count() - before
+
+        # Before the fix, each iteration leaked the handle duplicated for
+        # the child, so the count grew without bound.
+        self.assertLessEqual(leaked, 4)
+
+
 #
 # Issue 14151: Test invalid family on invalid environment
 #
diff --git a/Misc/NEWS.d/next/Library/2026-07-21-21-42-10.gh-issue-154208.Kq3vTd.rst b/Misc/NEWS.d/next/Library/2026-07-21-21-42-10.gh-issue-154208.Kq3vTd.rst
new file mode 100644
index 000000000000000..3dd2414b4144bac
--- /dev/null
+++ b/Misc/NEWS.d/next/Library/2026-07-21-21-42-10.gh-issue-154208.Kq3vTd.rst
@@ -0,0 +1,5 @@
+On Windows, :mod:`multiprocessing` no longer leaks a handle in the parent
+process for each handle passed to a child process which is terminated before
+it can steal them. Handles are now duplicated directly into the child
+process, so that they are released by the operating system if the child dies
+early.