- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,784
No secret developments - only a free model from the Internet. This is what scares researchers.
The researchers showed that malware with AI can already be collected from available models. The team created an experimental computer worm that attacks systems and spreads between devices. The prototype was run only in an isolated virtual environment, so it is not threatened with real computers.
The computer worm is different from a conventional malicious file by autonomous distribution. He does not need to wait for the user to download the attachment or manually launch the infected program. Once in the system, the worm is looking for new goals and tries to move on the network. Classic versions acted according to a predetermined scenario: the developer prescribed a set of steps, and the malware performed instructions.
AI is changing this scheme. The model inside the worm can adapt to the environment, look for weaknesses and choose a new path of attack, and not just repeat the same set of actions. University of Toronto professor David Lee, who is familiar with the work but did not participate in the study, considers the demonstration a wake-up call for the Uzbeks.
For the experiment, the researchers did not use the closed models of Anthropic or OpenAI. The team took a freely available AI model that can be downloaded from the Internet. This is an important detail: the risk is not limited to the largest developers and their internal security rules. Even with hard filters in commercial services, open models can get to people who are willing to use them for good purposes.
The work is published as a preprint and has not yet been reviewed. The authors believe that self-sustaining cyber threats on the basis of AI can no longer be considered an abstract danger. The experimental worm has shown the principle: the model can become part of malware, which itself seeks opportunities for further distribution.
The risk goes far beyond individual computers. Modern infrastructure depends on network systems in water supply, sewerage, energy, financial services, communications, healthcare, education, logistics, transport and the work of government agencies. If AI worms learn how to quickly find and exploit new vulnerabilities, not only corporate networks, but also critical infrastructure will be hit.
This technology also has a protective side. The same models can analyze the code, find vulnerabilities, check the settings and speed up the release of fixes. David Lee compares attack and defense with mirror processes: AI is able to amplify harmful tools, but the same methods can strengthen the defense.
The study shows that classical signatures and static detection rules will not be enough against malware that changes behavior as it spreads. Defenders need systems that can recognize not only the well-known code, but also the logic of the attack, which is rebuilt under a new environment.
The researchers showed that malware with AI can already be collected from available models. The team created an experimental computer worm that attacks systems and spreads between devices. The prototype was run only in an isolated virtual environment, so it is not threatened with real computers.
The computer worm is different from a conventional malicious file by autonomous distribution. He does not need to wait for the user to download the attachment or manually launch the infected program. Once in the system, the worm is looking for new goals and tries to move on the network. Classic versions acted according to a predetermined scenario: the developer prescribed a set of steps, and the malware performed instructions.
AI is changing this scheme. The model inside the worm can adapt to the environment, look for weaknesses and choose a new path of attack, and not just repeat the same set of actions. University of Toronto professor David Lee, who is familiar with the work but did not participate in the study, considers the demonstration a wake-up call for the Uzbeks.
For the experiment, the researchers did not use the closed models of Anthropic or OpenAI. The team took a freely available AI model that can be downloaded from the Internet. This is an important detail: the risk is not limited to the largest developers and their internal security rules. Even with hard filters in commercial services, open models can get to people who are willing to use them for good purposes.
The work is published as a preprint and has not yet been reviewed. The authors believe that self-sustaining cyber threats on the basis of AI can no longer be considered an abstract danger. The experimental worm has shown the principle: the model can become part of malware, which itself seeks opportunities for further distribution.
The risk goes far beyond individual computers. Modern infrastructure depends on network systems in water supply, sewerage, energy, financial services, communications, healthcare, education, logistics, transport and the work of government agencies. If AI worms learn how to quickly find and exploit new vulnerabilities, not only corporate networks, but also critical infrastructure will be hit.
This technology also has a protective side. The same models can analyze the code, find vulnerabilities, check the settings and speed up the release of fixes. David Lee compares attack and defense with mirror processes: AI is able to amplify harmful tools, but the same methods can strengthen the defense.
The study shows that classical signatures and static detection rules will not be enough against malware that changes behavior as it spreads. Defenders need systems that can recognize not only the well-known code, but also the logic of the attack, which is rebuilt under a new environment.
