diff --git a/rq/worker.py b/rq/worker.py
index 25b98e8..92e63b4 100644
--- a/rq/worker.py
+++ b/rq/worker.py
@@ -674,6 +674,9 @@ class Worker(object):
         either executes successfully or the status of the job is set to
         failed
         """
+
+        ret_val = None
+        job.started_at = job.started_at or utcnow()
         while True:
             try:
                 with UnixSignalDeathPenalty(self.job_monitoring_interval, HorseMonitorTimeoutException):
@@ -683,6 +686,12 @@ class Worker(object):
                 # Horse has not exited yet and is still running.
                 # Send a heartbeat to keep the worker alive.
                 self.heartbeat(self.job_monitoring_interval + 5)
+
+                # Kill the job from this side if something is really wrong (interpreter lock/etc).
+                if (utcnow() - job.started_at).total_seconds() > (job.timeout + 1):
+                    self.kill_horse()
+                    break
+
             except OSError as e:
                 # In case we encountered an OSError due to EINTR (which is
                 # caused by a SIGINT or SIGTERM signal during
diff --git a/tests/test_worker.py b/tests/test_worker.py
index e983bf5..7d8ed5a 100644
--- a/tests/test_worker.py
+++ b/tests/test_worker.py
@@ -1055,6 +1055,33 @@ class WorkerShutdownTestCase(TimeoutTestCase, RQTestCase):
         self.assertTrue(job in failed_job_registry)
         self.assertEqual(fooq.count, 0)
 
+    @slow
+    def test_work_horse_force_death(self):
+        """Simulate a frozen worker that doesn't observe the timeout properly.
+        Fake it by artificially setting the timeout of the parent process to
+        something much smaller after the process is already forked.
+        """
+        fooq = Queue('foo')
+        self.assertEqual(fooq.count, 0)
+        w = Worker(fooq)
+        sentinel_file = '/tmp/.rq_sentinel_work_horse_death'
+        if os.path.exists(sentinel_file):
+            os.remove(sentinel_file)
+        fooq.enqueue(create_file_after_timeout, sentinel_file, 100)
+        job, queue = w.dequeue_job_and_maintain_ttl(5)
+        w.fork_work_horse(job, queue)
+        job.timeout = 5
+        w.job_monitoring_interval = 1
+        now = utcnow()
+        w.monitor_work_horse(job)
+        fudge_factor = 1
+        total_time = w.job_monitoring_interval + 5 + fudge_factor
+        self.assertTrue((utcnow() - now).total_seconds() < total_time)
+        self.assertEqual(job.get_status(), JobStatus.FAILED)
+        failed_job_registry = FailedJobRegistry(queue=fooq)
+        self.assertTrue(job in failed_job_registry)
+        self.assertEqual(fooq.count, 0)
+
 
 def schedule_access_self():
     q = Queue('default', connection=get_current_connection())