AI security that assumes the attacker has root

An article discusses a concept in AI security that operates under the assumption that an attacker already has root access. The article is available to read and comments can be found on a corresponding discussion page. The topic has garnered some attention, with a point assigned to it.