AIs, including ones with Friendly goals, are apt to work to protect their goal systems from modification, as this will prevent their efforts from being directed towards things other than their (current) aims. There might be a window while the AI is mid-FOOM where it's vulnerable, but not a wide one.