Robot authorization guard blocks 109 hijacked actions in preliminary trials

Researchers report that an authorization system blocked all 109 unauthorized actions produced when attackers redirected AI planners in 240 live trials. The attacks sought to make robots depart from an approved mission while proposing actions that passed a physical safety check. The proposed AuthGuard-R system checks each executable action against a signed mission and limits tied to the robot, location, objects and time, alongside a separate safety check. The preliminary evaluation used two AI planners; a separate set of eleven hand-constructed protocol and policy attacks was also blocked.