Lepassaar says AI agents do not act independently — ERR News

In Estonia, Juhan Lepassaar, head of the EU Agency for Cybersecurity, said that fears of AI agents acting on their own and slipping out of human control are exaggerated. Risks arise when people configure such systems incorrectly, give them overly flexible instructions, or grant them excessive freedom of action, ERR News reports.

AI enhances the capabilities of malicious actors

Lepassaar explained that AI agents are programs that use large language models to plan, analyze and adjust actions until a goal is achieved. They are not conscious and do not act independently, but they may attempt to complete an assigned task outside the initial parameters.

According to him, modern models can significantly enhance hackers' capabilities. Systems trained on millions and billions of lines of open-source code are good at finding vulnerabilities and can help exploit them. At the same time, fixing such problems usually requires more time and human intervention.

More current news is available on the UA.News Telegram channel Telegram.

Cybercriminals also use AI for information manipulation and fraud. The technologies make it possible to create convincing synthetic voices and images to lure people into providing access codes, data or other information that can later be used to extort money.

Developers' responsibility

The agency head noted that Europe has a framework approach that classifies AI-related services and activities by risk level. It also defines the responsibility of developers and those implementing such systems, including for configuring and using the technologies to reduce risks.

Lepassaar added that the latest models can combine several minor vulnerabilities to penetrate a software system, although the basic methods of software testing are not new. He advised ordinary users not to share information with third parties and not to follow various links.