AI does not obey old cybersecurity rules, expert says

Traditional approaches to cybersecurity, based on strict restrictions, do not work with AI agents, said Yuri Maksimov, co-founder of the IT company Positive Technologies.

“Let’s say the agent won’t be able to do this, this and that. But the agent, he’s not a person, he won’t tolerate it. He just starts being stupid and not doing his job. That is, you can’t apply the old approaches of restrictions to AI,” he said at the plenary session of the Digital Solutions forum.

Therefore, as Maksimov noted, the task is formulated differently - instead of prohibitions, it is necessary to ensure that the agent is faithful to the person’s instructions. “That is, he did everything that was proportionate to what I instructed, and did not do anything that went beyond the reasonable boundaries of all this,” he explained.

At the same time, according to the specialist, it is impossible to solve this problem without artificial intelligence itself - a person will not be able to analyze everything that the agent does. Therefore, according to him, a scheme of several AIs is created: one system, for example, builds a virtual bank, the second tries to compromise it, the third learns to detect an attack. All of them are closed in a chain and are constantly being improved, and at the head of the entire structure is a person.

Separately, he touched upon the issue of responsibility. According to him, when a person entrusts a task to an agent and the agent does something wrong, the person can abdicate responsibility. However, this approach will not suit either the state or business, the expert pointed out.