Model has 'improvements in honesty and reductions in deceptive behaviour' – after AI went rogue in July ...