Weight theft scenarios focus on external hackers, but insiders with legitimate access are the harder problem. What insider-threat controls actually exist at frontier labs, and are they proportionate to the stakes?