Shocking warning: Artificial intelligence may erupt this year, take control and spread into unexpected areas

Washington - Ofoq News - Anthropic, an American company specializing in artificial intelligence, warned in a new study published this week that small artificial intelligence systems may, by this year, acquire the ability to self-replicate, spread among devices connected to the Internet, and continue to work without the need for a human operator at any stage.

Researchers have been able to demonstrate important parts of this process, and they estimate that the remaining technology gap until reliable, independent deployment is rapidly closing.

This is a very important warning. If these systems are combined with computer hacking capabilities, they may constitute a new type of threat: software that does not content itself with a single intrusion, but rather seeks to understand its environment, find additional devices, copy itself, and continue operating even after its previous attempts have been repelled. This does not prove that such a system is spreading freely on the Internet, but rather it is a warning of a capability that may develop this year.

This has direct implications for everyone who owns a mobile phone. Imagine a malicious system that manages to compromise a device through a security flaw or by stealing access permissions. This system reads messages, copies photos and documents, collects login data, and attempts to move to other devices. He may later try to erase traces of his activity and continue working elsewhere. The phone owner may not even notice what happened.

This opens up a wide world of dangerous possibilities. The phone itself may sometimes be used as a gateway to access your bank account, email, workplace, medical records, and government systems. If a hostile system can gain the necessary powers through it, it may attempt to penetrate sensitive databases and change not only the information we read, but also the information on which decisions that affect our lives are made.

Suppose, for example, a student took a university exam and scored 62. An AI system, if it can hack into a grade database and gain the necessary editing permissions, might attempt to change the grade to a 96. It might also do the opposite: convert a perfect grade to a failing grade, change attendance records, or disable credential eligibility data. If he also managed to destroy the documentation of the changes and the control mechanisms, a troubling question arises: how can the true degree be proven?

This example seems funny at first glance, until it is applied in a hospital. A compromised system with access to a medical record might attempt to change test results, delete information about drug allergies, or tamper with vital details in a patient's file. The doctor, based on incorrect data, may make a dangerous decision. Even with protection and backup mechanisms in place, the mere possibility of compromising the integrity of medical information poses a serious threat.

In the banking system, the consequences could be dire. A system that gains unauthorized access may attempt to alter account data, forge payment instructions, transfer funds, and disrupt financial records. Banks have sophisticated control and security systems, but the scenario of automated and persistent attacks requires them to deal with an adversary capable of trying different courses of action at speed and scale.

The biggest risk is not only losing money, but losing trust in the information. What if a person cannot verify the correct balance in his bank account? What if the patient cannot confirm the accuracy of his medical record? What if the student had to prove that the grade recorded on the computer did not represent the grade he actually obtained?

This is the tip of the iceberg. Land registries, population registries, tax records, insurance systems, pension payments and security databases all depend on the reliability of data. A hostile regime can alter sensitive information, potentially harming civil rights, property, economic activity, and the functioning of state institutions.

This concern also extends to critical infrastructure. Electricity, water, communications, and transportation systems rely heavily on computers and data transmission networks. Hacking sensitive operating systems is no easy feat, and many of them are isolated and protected. However, an automated system capable of continuing to search for vulnerabilities and try new hacking methods may increase the burden on defense systems and create additional risks.

The novelty of Anthropic's research is that such systems can dispense with huge computing centers. The researchers were able to run a model containing 26 billion parameters using 15 inexpensive Android TV receivers, each priced at about US$55. The model was shared between machines, each of which performed part of the computation. The system was slow, but its operation itself demonstrated how relatively simple hardware could be harnessed for AI tasks.

In another experiment, a small model was able to work on eight out of nine types of devices tested. As for the more powerful models, they were able to perform a series of operations in simplified test environments, which included finding another computer, transferring the system to it, and running it. Smaller models still have difficulty completing the entire process reliably, but researchers expect this capability to mature as early as this year.

Here lies the most dangerous aspect of this research. An ordinary computer program performs pre-defined tasks, while an advanced artificial intelligence system can analyze the situation, draw conclusions, choose alternative paths, and attempt to overcome obstacles. If it were also equipped with the ability to self-replicate, it would become a program that spreads spontaneously and adapts to changing conditions.

This is not necessarily a rebellion by a machine that has acquired consciousness or feelings. It is enough for the system to act contrary to the intentions of its operators, to continue performing a task that they were unable to stop, and to move between devices without obtaining human approval for each action. This is the practical and disturbing meaning of losing control.

This system may try to hide its tracks, such as deleting records, using stolen digital identities, and changing its operation to avoid detection. This does not mean that it will disappear completely, as independent monitoring systems, backups, and system logs may reveal its activities. But as its ability to adapt improves, it may become more difficult to detect it in time.

The problem is exacerbated when the system is not installed on a single computer. If copies of it are running on multiple machines, disabling one server will not necessarily stop the other servers from working. It may require coordinated action from technology companies, telecommunications service providers and authorities in various countries to stop its spread.

The risk is not limited to one system. As small models become more powerful and less expensive, criminal organizations, adversarial actors, and even individuals may gain access to capabilities that previously required significant resources. The world may face multiple automated attacks operating in parallel, searching for vulnerabilities and changing their operating methods.

It is important to clarify that the study does not prove that AI systems are actually capable of implementing all of these scenarios. Self-reproduction, hacking into banking databases, and manipulating medical data are all different capabilities, and moving between them requires bypassing layers of protection and authorization. However, the warning indicates the approach of a very dangerous technological element: the ability to continue spreading and operating without constant human oversight.

The digital world is based on trust. We believe that money in banks is accurately recorded, that our medical examinations are correct, that our university degrees reflect our achievements, and that the state maintains reliable records of its citizens. If hostile systems can operate independently against these databases, the conflict will not only be about protecting the computers, but also about the authenticity of the data recorded in them.

So, Anthropic's warning is not just a long-term prediction about robots taking over the world one day, but rather indicates a technological development that may be completed this year, and requires immediate preparation from governments, banks, hospitals, universities and technology companies.

The troubling question is not just whether an AI system will be able to hack our computers, but what would happen if it could copy itself to other computers, continue to operate even when we try to stop it, and change information that we have become accustomed to taking as absolute truth.

On this day, the biggest fear will not only be what AI might do in the future, but what it has already done without our knowledge, and whether we will be able to distinguish between the original information and the digital reality it has changed.