Prime Intellect says an intruder replaced a hosted model A November notice reports unauthorized model weights behind one inference endpoint. Similar capability scores concealed changed refusal behavior; independent tool permissions still matter.