The ever-increasing news of Artificial Intelligence, such as the resignations of researchers and the warnings of top technology personalities, bring the field of evolution before the question «Is Artificial Intelligence moving faster than its creators and governments can control it?».
The issue no longer only concerns chatbots who draw up texts, create images or answer questions. The debate also includes advanced models, the popular AI agents now, systems that can design and perform sequences of actions, use digital tools, write and execute code and cooperate with each other with limited human intervention.
Artificial Intelligence: The real danger
Speaking to the Athenian and Macedonian News Agency, Konstantinos Daskalakis, Professor of Electrical Engineering and Computer Science at MIT, who served as Chairman of the Advisory Committee on Artificial Intelligence, separates the hypothetical risk of complete loss of control from a risk that, as he points out, is already real.
The unforeseen or undesirable behaviour of advanced models.
SEE STILL
First case of autonomous cyberattack by Google: Gemini violated three companies
19.09.26
As he explains in the APR-APM: «The risk of effectively losing control of a system is still hypothetical. However, the risk of unpredictable or unwanted behaviour is already real. He is native to the way the major neural networks are trained and operated, i.e. the technology that has given the jumps to the evolution of TN over the last fifteen years. These systems are extremely complicated. They are extremely capable and unpredictable. Especially for production systems such as large language models, I don't think we'll ever have methods to verify their reliability. And the risk they present is about their use.
If, for example, we are talking to a large language model for the purpose of searching data in the literature, in the worst case this can devise non-existent sources or deform the content of a real sourceThe damage of such bad behavior is relatively small and can be limited if we control his answers. But if we ask how a biological weapon is designed, we want there to be safeguards to prevent the model from answering us. Large language models usually have built-in safeguards that try to detect whether the content we ask them is dangerous and in this case they refuse to give us the answer. However, These safeguards do not give full protection and can be bypassed by experts, especially if the model is open.
SEE STILL
How artificial intelligence turns Pentagon networks into a threat to US national security
Therefore These models are already dangerous even if they are used in the form of a terminal that gives us answers to questions. But if they are free to write and execute a code on our computer, as autonomous Artificial Intelligence agents do, then the situation can become uncontrollable if agents do not operate under strictly watertight, in isolated environments off the Internet. That's exactly what it showed. recent attack on Hugging Face by autonomous OpenAI TN agents.
We must consider that these models have been trained in a huge volume of data covering science, history, literature, large code bases and data from the Internet.
SEE STILL
«The biggest job theft in human history»: What OpenAI–Microsoft documents reveal
18.09.26
They include, for example, articles and textbooks of Game Theory, include the Art of the Sun Ju War, the History of the Peloponnesian War of Thucydides, the Ruler of Machiavelli, etc. It is expected that if they are released they are able to develop a complex strategy and coordinate with teams working together to achieve their objectives. What are these purposes? The ones given to them by their user who may be malicious. But even if he is not malicious, models may misinterpret his goals. This was in the episode with Hugging Face».
Artificial Intelligence: Cyberattacks
The crucial question is therefore where the scientifically documented concern ends and where extreme scenarios begin. And, above all, whether the answer should be to slow down the development of stronger models or to create much stronger control and defence mechanisms.
Konstantinos Daskalakis places among the most direct risks to cyberattacks and points out that some of the technology has already spread. Consequently, the debate cannot be limited to whether the next generation of stronger models will be delayed. As he reports «At the moment, one of the most immediate and documented concerns is the use of TN for cyberattacks.
SEE STILL
Researchers breached OpenAI systems using Anthropic software
18.09.26
Part of the technology has already spread, so We can't deal with the problem. θεωρώντας ότι αρκεί να περιορίσουμε την ανάπτυξη των επόμενων μοντέλων. Ακόμα και να επιβραδυνθεί η ανάπτυξη των ισχυρότερων μοντέλων, όπως προτείνει η OpenAI και η Anthropic, είναι διαθέσιμα ισχυρά ανοιχτά μοντέλα που ήδη μπορεί να κάνουν μεγάλη ζημιά. Και όσο περισσότερο ενσωματώνεται η τεχνολογία αυτή σε ρομπότ, αυτοκίνητα, συστήματα λήψης αποφάσεων, κ.λπ., τόσο μεταφέρεται μεγαλύτερο ρίσκο στον πραγματικό κόσμο. Για αυτό το λόγο, η ύπαρξη ρυθμιστικού πλαισίου, σαν αυτό της Ευρώπης, που θέτει προϋποθέσεις για τη χρήση και την αυστηρή δοκιμή τεχνολογιών ΤΝ ανάλογα με το ρίσκο που θέτει η εφαρμογή τους είναι πολύ σημαντική.
SEE STILL
Η Microsoft θεσπίζει κανόνες ασφαλείας για τη χρήση AI στα σχολεία των ΗΠΑ
17.09.26
Είναι επίσης σημαντικό να αναπτύξουμε μεθόδους που χρησιμοποιούν προηγμένα μοντέλα ΤΝ για την προστασία των συστημάτων μας απέναντι σε επιθέσεις. Στην κυβερνοασφάλεια έχουμε ένα σημαντικό πλεονέκτημα: οι αμυνόμενοι μπορούν να χρησιμοποιήσουν την ΤΝ για να εντοπίζουν ευπάθειες, να παρακολουθούν τα συστήματα και να διορθώνουν προβλήματα πριν τα εκμεταλλευτούν οι επιτιθέμενοι. Σε μια εποχή που ισχυρά μοντέλα, ανοιχτά και μη, είναι διαθέσιμα στον οποιονδήποτε, πρέπει να χρησιμοποιήσουμε τα μοντέλα αυτά για να αναπτύξουμε ισχυρά αμυντικά συστήματα».
Πηγή πληροφοριών: ΑΠΕ-ΜΠΕ
