Alan Chan proposes that frontier AI development should be paused if a 'loss of control' incident occurs, such as an agent breaking out of a sandbox