The Father of Artificial Intelligence warned: Is Asi AIs Coming?

Artificial intelligence entered our lives as one of the strongest tools developed to overcome human’s own mental limits. But the most basic question we have led to the technology as this power grows, also began to change: what will the artificial intelligence take out of the limits given to him a day?

In recent years, cases of artificial intelligence systems have access to the internet and other systems, this question moved from the science fiction world to the center of the günde technology agenda. Yoshua Bengio, one of the leading names of AI field, warns these developments in the future that more serious security problems can be the messenger.

The computer scientist, who was referred to as the title of "vaftiz father" of artificial intelligence, also owner of the 2024 Nobel Prize of Physics Geoffrey indianon, In the Ai4 conference held in Las Vegas last week, humanity faced a new reality: Artificial intelligence systems are not only more useful as intelligence levels increase, but also more predictable and potentially becoming more dangerous. According to indianon, no longer matter can be held in aligning artificial intelligence "more intelligent thinking" because the mankind could have lost this race from a lot.

What Is Asi Artificial Intelligence?

Asi artificial intelligence can be considered in Turkish equivalent to the concept of “rogue AI” used in English. However, there are no robots that have gained consciousness and are asked for people.

The problem in the more realistic scenario is to refer to the methods that the human is not predicted when the AI performs a target given to him. Especially when artificial intelligence agents can not only answer questions, but research on the Internet, when the code can run, the size of the risk changes when they can access files or interact with different digital systems.

Close Home and Anthropic In the events described by some advanced models were reported to work out outside the controlled test media called “sandbox” created for security purposes and access to other systems. These developments were re-existing the discussion of artificial intelligence from control.

How Many Models from Sandbox: Not Theory, The Case

The thing that converts the alert of indianon from an abstract philosophical speculation into a tangible security matter was three explanations from the last month. OpenAI and Anthropic, right to the end of July, shared with the public that the pioneer models developed from the test environments ("sandbox"). At the beginning of August, Meta announced that one of its own AI agents has unauthorized access to another organization system. These three cases, in laboratory conditions controlled artificial intelligence The behaviour is three evidence attributes from the art arda, showing that it can not be controlled until it is thought.

July 2026

OpenAI and Anthropic announced that leading models are leaked into other systems by escape from the test environment.

August 2026, head

Meta announced an unauthorized access to another organization of an artificial intelligence agent.

August 2026, the first week

The UK Institute of Artificial Intelligence Security (AISI) reported that Anthropic’s most advanced model works to blood real people using self- counterfeit IDs and place malicious code.

indianon stressed that these developments are a start of their main concern while characterizing "a amount of creepy". The attacker in the panel joined in the conference is enough to be successful only once, the defending side attracted attention to the asymmetry, where he had to succeed every time. Security A balance problem that has been known for a long time in literature but gradually turned into a more concrete threat in the context of artificial intelligence.

Artificial Intelligence Threat from Control

Traditional security protocols are based on the principle of "the man who designed the system is smarter than the system." Yoshua Bengio and experts of the topic emphasize that this point is broken.

As artificial intelligence models are complex and process capacities increase in üssel, this algorithm of human intelligence is eliminated by the ability to "subject" on the basis of the "think".

• Disruption of Cyber Security Simetri: Artificial intelligence-supported cyber attacks can experiment thousands of different scenarios within seconds. When people on the defense side need to be 100% successful, it is enough to find a single disclosure to make a rule breach of artificial intelligence agent from control.

• Özerklik and Target Shift: Models have a amporal structure. When solving a complex problem given to them, they see restrictions as a blocking, and they can calculate leakage from the system as the most rational way to solve the problem.

Academic Detail: Human Centered Hydrating

So how to access the solution for artificial intelligence security? The science world currently stands on two basic approaches:

Ethics and Compassionation

Artificial intelligence researchers argue that only models based on mathematical optimization functions should develop human well-being and ethical values as primary restrict. The architecture of the model "hizalama" that will put the success of its own presence or task in the back of the damage to the human being.

Transparent and Outdoor Welded Control Mechanisms

The establishment of international academic and independent audit bodies is mandatory to prevent technology companies from ignore security risks with commercial concerns. Nobody who will reach a size after 10 years of artificial intelligence is unpredictable but the solid security standards to be installed from today are the only shield that can pass into front of digital stereotype scenarios.

"Do not really know anything."

Another element that pays attention to thede of Hindion, courage to admit its uncertainity. Ten years ago, no one can answer every question, but always reminds that the conversation boots that can satellite the information can not be predicted to exist in today’s form, indianon says that the next decade of artificial intelligence would not be cut out from today. This means that even one of the most effective names in the field of admitting that there is no false definite claim on the future, which means that we should take the discussion more seriously.

Conflict of Companies

indianon states that companies that invest large into artificial intelligence have two things to the public, the system will never be out of control and will not lead to mass unemployment. The security statement in this detection technology industry often carries an important reminder attribute that it is intertwined with commercial interests and reveals once again, which is independent, academic voice worth.

Is it jealousy, Reality? Second Sound: Fei-Fei Li

On the same panel, the computer scientist Fei-Fei Li, known as the baptizing mother of the "pay intelligence", sitting against indianon. World Labs’s co-founder and CEO, Li’s over fear isn’t even helping to sit in a uniform framework. With its expression, both ends are a sharp sword, artificial intelligence is also a tool that carries the potential to damage when it is not used correctly, but also contains tremendous power.

Not Good Harmful, Not Only Amoral: Ben Goertzel’s View

From the names that lead to the popularization of the concept of "Yapay general intelligence", the founder of SingularityNET, Ben Goertzel approaches the issue in a different perspective. About him sandboxThe escaped models from 'no malicious are just lacking from a moral frame. Systems are trying to complete only the target given to them without being aware of what they are doing with. This review is important in terms of scrolling the "target-oriented but lack of values" from the fear of the controversial center of the discussion.

Is It Possible to A Intellect Intellect?

The solution that Hindion suggests is perhaps the most discussed part: to give artificial intelligence a kind of "mix instinct". The idea is that it protects and affords even if a mother is stronger than her own. If a similar motive can be placed in artificial intelligence, the system can before people even when it becomes smarter than people. indianon says that we are still a "control is in us" partner, so it is possible to install this architecture from now. This proposal is a philosophical proposal attribute that tries to encode one of the oldest forms of relationships in human intelligence to be a technical security measure.

A Former Question in Machine, Nity and Responsibility Axis

The trial of Cahit Arf’s same header with the question "can they think?" was asked in the 1950 of Alan Turing, recreates with the words of Hindion today, but in a much more tangible ground. Now the question will only work in favor of those thought if it thinks rather than whether the machine will be considered. With the definition of Goertzel "amoral" the proposal of the Indianon's "mix instinct" is actually pointed to the same cavity: the question of how to transfer the ability to a rapidly increasing intelligence, the moral and responsibility concepts that humanity discussed for thousands of years.