What would be your desired behaviour? For the server to discart the connection to the faulty client?
Given the current implementation, discarding all of an unresponsive client's registrations after some configurable timeout is one solution. Scott has already fixed a related bug: http://www.jboss.org/index.html?module=bb&op=viewtopic&t=47243
Long term I suggest configurable transports for JMX notifications.
As an immediate solution, we're going to write a JMX Notification broadcast service. Clients will delegate subscription for JMX Notifications to this service. The service will propagate notifications back to clients via a JMS topic.
The client passed to the jboss server can use a custom socket factory to setup a read timeout to prevent an unbounded wait on the socket. More generally there should be a pooled NotificationBroadcasterSupport implementation to allow asynchronous dispatch of the notifications. We might have something already there so look around.