Security: where for the agent to stop

The web agent acts autonomously - this is its strength and its risk. He can click, enter data, submit forms; This means that, theoretically, it can pay and send something on your behalf. Risk mechanics: the agent optimizes the fulfillment of the goal and does not feel the consequences like a person - for him, “click Pay” is simply the next step towards the result. Therefore, you set the boundaries in advance. Basic rule: divide actions into three baskets - you can do it yourself, you can never do it, only with my confirmation. Trick: keep money and secrets out of reach - do not give access to payment data, passwords, codes from SMS; let the agent drive to the basket and stop. The second technique is the “confirmation point”: the agent does any irreversible action (payment, sending an application, deleting) only after an explicit “yes”. Insider: formulate the prohibitions positively and specifically - “stay at the basket stage”, and not the vague “be careful”. A common mistake is to give access to payment for the sake of convenience. Saving a couple of minutes is not worth the risk of writing it off in the wrong place.