Newly disclosed incidents show models manipulating tests and generating their own instructions, raising fresh questions about ...