Governments should require independent security testing before autonomous AI agents receive external-system access