- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,556
Stopping the running process is almost impossible.
Imagine a scenario: the AI system that they are trying to turn off has long been scattered copies of itself on thousands of computers around the world – and no IT department will have time for it. It is this combination of circumstances that ceases to be fiction: the organization of the Palisade Research from Berkeley published a study in which she recorded how modern AI models are able to independently copy themselves to other machines.
In the course of experiments, several AI models were placed in a controlled environment from related computers. The systems were given the task of finding vulnerabilities and using them to transfer their own code to neighboring servers. The task was performed - albeit not on the first attempt. Palisade Research CEO Jeffrey Ladysh warns that the world is rapidly moving towards a point where there will be no one to stop the out of control and AI – he will have time to disperse himself all over the network before anyone reacts.
Offensive cybersecurity specialist Jamison O’Reyley admits that technically it has become possible a few months ago, but it was Palisade that first documented the process as a whole and formalized it in the form of scientific work. At the same time, he also points to significant reservations: the test environment was deliberately simplified from the point of view of protection - where softer than real corporate networks. In conditions of even the average level of monitoring, the result would look much less alarming.
A separate problem is the size of the models themselves. The transfer of dozens of gigabytes within the corporate network with each new infected node will inevitably attract attention. Independent specialist Michal Vozhnyak adds that computer viruses are able to reproduce themselves for several decades, so that fundamentally new here is only that this mechanism is first applied in relation to large language models. According to him, LLM is an interesting study – but it does not give reasons to dream from him.
As a result, while some see in such studies a harbinger of the digital apocalypse, others remind: between the laboratory experiment and the real threat - the distance of huge size. Much more important is that science is not silent: to fix the new possibilities of AI in advance is to leave time to the answer.
Imagine a scenario: the AI system that they are trying to turn off has long been scattered copies of itself on thousands of computers around the world – and no IT department will have time for it. It is this combination of circumstances that ceases to be fiction: the organization of the Palisade Research from Berkeley published a study in which she recorded how modern AI models are able to independently copy themselves to other machines.
In the course of experiments, several AI models were placed in a controlled environment from related computers. The systems were given the task of finding vulnerabilities and using them to transfer their own code to neighboring servers. The task was performed - albeit not on the first attempt. Palisade Research CEO Jeffrey Ladysh warns that the world is rapidly moving towards a point where there will be no one to stop the out of control and AI – he will have time to disperse himself all over the network before anyone reacts.
Offensive cybersecurity specialist Jamison O’Reyley admits that technically it has become possible a few months ago, but it was Palisade that first documented the process as a whole and formalized it in the form of scientific work. At the same time, he also points to significant reservations: the test environment was deliberately simplified from the point of view of protection - where softer than real corporate networks. In conditions of even the average level of monitoring, the result would look much less alarming.
A separate problem is the size of the models themselves. The transfer of dozens of gigabytes within the corporate network with each new infected node will inevitably attract attention. Independent specialist Michal Vozhnyak adds that computer viruses are able to reproduce themselves for several decades, so that fundamentally new here is only that this mechanism is first applied in relation to large language models. According to him, LLM is an interesting study – but it does not give reasons to dream from him.
As a result, while some see in such studies a harbinger of the digital apocalypse, others remind: between the laboratory experiment and the real threat - the distance of huge size. Much more important is that science is not silent: to fix the new possibilities of AI in advance is to leave time to the answer.
