Page 165 - CyberSecurityEssentialsEnglish
P. 165

4.4 Delegate: Give the System a Role, Not Unlimited Authority

           ◆ Key idea: delegation is a matter of degree.

           AI can suggest, draft, analyze, recommend, or execute. Every move to a higher degree
           increases the need for constraints, limited permissions, human approval, and monitoring.
           A task that is safe in draft mode may become dangerous if the system is given permission
           to send, delete, or purchase.

         Good delegation begins by breaking the task into parts: what requires speed and a pattern

         the system can support, and what requires responsibility, expertise, or a human relationship
         and  must  stay  with  a  person.  In  preparing  a  response  to  a  complaint,  the  system  can

         summarize  the  problem  and  propose  wording,  while  the  employee  decides  the
         compensation and the final tone and checks the customer’s data.

         Automation bias is the tendency to accept the system’s suggestion because it looks neutral
         or technical. It appears especially under time pressure or with repeated tasks. The remedy

         is not to write “review carefully” but to design a specific review point:  which fields are
         checked? Which cases are excluded? And who approves?
         Human-in-the-loop  means  that  an  important  action  is  not  completed  before  a  person

         intervenes.  Human-on-the-loop  means  that  the  system  operates  within  limits  under
         monitoring, with the ability to stop it. The choice between them depends on the speed of
         the event, the impact of error, and the possibility of reversal.


         Reduce permissions according to the principle of least privilege. If the system needs to read

         one file, do not give it access to a whole folder. If the requirement is to prepare a draft email,
         do not give it automatic sending. And if it needs to create a calendar event, make the final

         step a presentation for approval.

         When AI is connected to tools or databases, outputs become actions. Here thinking must be
         separated  from  execution:  the  system  produces  a  plan  or  a  proposed  command,  an
         independent layer checks the conditions and permissions, and then a clear confirmation is
         requested before any irreversible step.

























        164
   160   161   162   163   164   165   166   167   168   169   170