Skip to main content
Advertisement
Live broadcast
Main slide
Beginning of the article
Озвучить текст
Select important
On
Off

Researchers at the American company Anthropic have admitted that by the end of 2026, small artificial intelligence models may approach the ability to independently spread between devices connected to the Internet. Scientists described the potential threat as "digital mold" — a distributed system that uses other people's computing resources without the knowledge of their owners. Household appliances, such as televisions, refrigerators, and robot vacuum cleaners, are considered as a possible distribution medium. However, experts interviewed by Izvestia believe that the transformation of such systems into a massive threat is still unlikely. What is a "digital mold", how it can be dangerous and how to protect home appliances — in the Izvestia material.

What is a "digital mold" and how does it work?

The term "digital mold" is used by the Anthropic researchers to describe a hypothetical autonomous AI system capable of spreading between devices, creating its own copies and using available computing power. Unlike traditional malware, such a system will potentially be able to adapt to the hardware features and continue working without constant human commands.

As Andrey Kuznetsov, director of machine learning at Positive Technologies, explained to Izvestia, we are talking about one of the scenarios considered by the researchers, and not about the confirmed appearance of a new type of cyber threat. The main objective of the experiments was to test the ability of small neural networks to independently copy themselves to other devices and run on them.

— The difference from the usual botnets is not in the method of infection. Most botnets are controlled from a center that can be calculated and disabled, and work according to pre-written instructions for a specific type of device. An AI worm does not need such a center: the model on each device decides for itself what to do next and can sort out unfamiliar equipment," the expert said.

During the experiments, a small model was able to independently configure the launch of another neural network on eight of the nine types of equipment, including single-board computers and Mac devices. The researchers also found that task performance increases when information is transferred between successive instances of the model.

If the new copy received notes from the previous one about failed attempts, the success rate increased from the initial 45% to 61-84%. This shows that such systems are potentially able to use the accumulated experience to adapt to unfamiliar equipment.

Traditional network worms are usually created for specific platforms, whereas the AI model can independently determine the appropriate launch method, explained Dmitry Kryukov, head of the MTS Link AI department.

"The AI worm carries the brains with it: the weights of the model are included in the payload, and on an unfamiliar device it figures out how to launch itself," the specialist noted.

At the same time, Vladislav Tushkanov, head of the Kaspersky Lab machine learning technology research group, suggested considering the "digital mold" in the context of earlier thought experiments about the uncontrolled development of technology.

In particular, he recalled the concept of "gray slime" proposed by engineer Eric Drexler. According to this scenario, self-replicating nanomachines are theoretically capable of using the surrounding matter to create their own copies, gradually absorbing available resources.

"Digital mold," according to Tushkanov, continues this idea. If language models can be reproduced on Internet-connected equipment, they can theoretically spread around the world, and removing them completely will be a difficult task.

Why household appliances can become a medium for the spread of AI

One of the prerequisites for the emergence of such systems was the development of compact neural networks, which require fewer and fewer computing resources. At the same time, the number of household devices constantly connected to the Internet is increasing, from televisions and video surveillance cameras to smart speakers and refrigerators.

According to Vladislav Tushkanov, modern models are indeed becoming more efficient, and their compression methods allow complex algorithms to be run on less powerful hardware. According to an analysis by the Epoch AI research organization, cited by the specialist, about six months pass between the appearance of an advanced model and the possibility of launching it on consumer video cards.

However, this does not mean that modern neural networks are already able to fully operate on ordinary household appliances.

— When we talk about "consumer equipment", we are talking about powerful video cards or devices with shared video and RAM for hundreds of thousands or even millions of rubles, and not about refrigerators or even ordinary computers, - the expert drew attention.

According to him, the probability that compact models suitable for running on a smart TV will acquire the capabilities of advanced neural networks in the near future remains quite low with existing architectures.

Anthropic's experiments also revealed significant limitations of the technology. According to Andrey Kuznetsov, the researchers manually launched the model on a bundle of 15 inexpensive set-top boxes. However, such a system worked hundreds of times slower than the usual chatbots.

According to the calculations of the authors of the work, one copy to a new device in such conditions could take from one and a half days to several days. However, even combining a large number of weak devices does not guarantee sufficient performance.

As an example, the researchers examined the Kimwolf botnet, which unites more than 2 million infected devices. According to Kuznetsov's estimate, the computing capabilities of a network of this scale would not exceed the performance of 40 server-side AI accelerators.

In addition, the distribution process was significantly simplified in the laboratory: devices on the test network were not password-protected, and direct hacking of the equipment was not checked. The full cycle of self-copying was demonstrated only with the participation of large models that cannot be placed on ordinary household appliances.

To organize a distributed launch, the researchers needed the powerful GLM-5.2 model, Vladislav Tushkanov clarified, while the much more compact Llama 3 8B was transferred to less productive equipment. The latter by itself did not have the necessary capabilities for further independent distribution.

Nikita Vekesser, head of Nova AI and StarGuard AI products at Orion soft, also believes that the computing resources of typical household appliances are not enough for even small language models to work independently. However, such devices can potentially be used to transfer information, store system components, or perform simple computing tasks.

At the same time, the spread of malware is facilitated not so much by the power of the equipment as by its insufficient security. Many devices are constantly connected to the Network, rarely receive updates, and use standard security settings. Surveillance cameras, routers, and media set-top boxes remain the most vulnerable, according to Wekesser. They work almost continuously on the Internet, and their software is often updated much less frequently than the operating systems of computers and smartphones.

How "digital mold" can threaten users

The potential proliferation of autonomous AI systems raises concerns primarily because of the possibility of unauthorized use of computing resources. An infected device can perform extraneous tasks, transfer large amounts of data, and place additional strain on the processor and memory.

For the owner, such activity is likely to manifest itself in slowing down the equipment, heating it up and constantly exchanging information with other network nodes, said Andrey Kuznetsov.

At the same time, the expert does not expect a significant increase in energy costs. Much more serious is the possibility of using infected equipment to infiltrate the home network.

—Cameras, microphones, and personal data are under threat not because of AI, but because of the hacking itself: attackers still get such access today, without any AI," the expert explained.

Dmitry Kryukov holds a similar position. He noted that the Anthropic study was not about stealing personal data or gaining access to surveillance cameras. However, if an attacker manages to run their own code on the device, all the data and functions that this code accesses are potentially at risk.

An additional danger is the possibility of forming distributed networks of infected equipment. Such an infrastructure is already being used by cybercriminals to carry out DDoS attacks, organize proxy services, and other illegal activities.

The more urgent threat, according to Vladislav Tushkanov, is not related to self-replicating compact models, but to the use of powerful language models to automate cyber attacks. Attackers can use AI agents to search for vulnerabilities in corporate networks, analyze program code, and prepare attacks. The development of such tools will potentially significantly expand the scale of criminal activity.

— If earlier the resources of cyber intruders were limited, now they can scale their activities and attack targets that previously would not have been worth the time and resources spent, — said the specialist.

Andrey Kuznetsov drew attention to another scenario — the spread of autonomous AI worms inside the corporate infrastructure. Today, one of the most extreme ways to stop a large-scale cyberattack remains the emergency isolation of the company from external networks. This approach allows you to disrupt the attacker's communication with infected systems.

— The worm, which works on the basis of an AI model and has already captured part of the victim's computing power, will continue to spread inside the data center and the entire infrastructure after such a shutdown. This is a new round of complication of protection," the expert warned.

Is it possible to stop the spread of "digital mold"

Despite the potential risks, experts interviewed by Izvestia believe that existing cybersecurity mechanisms are capable of countering a significant part of such threats.

According to Dmitry Kryukov, detecting an autonomous AI worm may be easier than expected. To launch the language model, it is necessary to transfer significant amounts of data between devices, and its operation is accompanied by characteristic network activity.

— The model weighs gigabytes, and it needs to be transferred to each new device. In addition, the console starts constantly driving traffic with neighbors with whom it was previously silent. The provider sees this," the specialist explained.

Monitoring network traffic and isolating infected equipment can be the most effective counteraction tools. Traditional antivirus programs have limitations. They cannot be installed on most household devices, and the software that ensures the operation of the neural network does not necessarily contain signs of malicious code by itself.

Kryukov gave an example of fighting the Mirai botnet, when Internet service providers isolated infected subscribers and sent them instructions on cleaning the equipment. In two weeks, 92% of infections were eliminated.

This problem is of particular relevance for the Russian low-cost electronics market. According to Andrey Kuznetsov, Russia is among the countries with a high concentration of devices infected with the Kimwolf botnet. Inexpensive Android consoles, which can have debugging mode enabled by default, play a significant role in the spread of malware.

To reduce risks, experts recommend avoiding standard passwords, installing updates in a timely manner, disabling unnecessary remote access functions, and placing smart devices on a separate guest Wi-Fi network.

Specialists assign a separate responsibility to equipment manufacturers. According to Dmitry Kryukov, devices should be supplied with unique passwords, debugging modes disabled by default, and the ability to regularly update software.

At the same time, the development of distributed computing is not necessarily associated with malicious activity. As the expert noted, similar technologies can be used to voluntarily pool users' computing power.

As an example, he cited the Folding@home project, which used the resources of volunteer computers for scientific calculations. In the future, a similar approach can be used to perform certain tasks of artificial intelligence.

The fundamental difference between a legal system and a malicious one is the consent of the equipment owner, transparent conditions for using its resources, and reliable protection of personal data.

Переведено сервисом «Яндекс Переводчик»

Live broadcast