Artificial intelligence has many advantages that help many users to be more productive, while it also has enough use during the generation of certain content. But in addition to everything related to technology, in addition to being used for purposes linked to its original use, it can also be used to cause problems To other people, and now some researchers have managed to obtain part of the IA more famous, they directly believe malware.
Every day, companies that are at the origin of the different models that allow you to use an AI (LLM) are trying to establish a series of filters to control the capacities that this software has since in the event that the limits were not established, they could obviously generate or create extremely dangerous content. But the eldest issue that these limitations have is that there are ways to ensure that its own intelligence artificial It is ignored, something that has been demonstrated several times and which continues today to operate in some cases despite the efforts of large companies behind them to prevent it from happening.
The big problem of AI is really likely to “jailbreak”
Artificial intelligences currently have a series of ways to generate content, some work as if they were an assistant who responds to requests made by the user, which allows you to maintain a conversation with them when they can create from zero, even whole lines of code. This can greatly facilitate work for example to a developer, because with a single question, a chatbot can generate the code to implement a function in a game.
But the fact that they can do so implies that it will not always be used for good, because they can also create a malicious code jumping all the restrictions that companies have implemented behind their development. It is something that the researchers of Cato Ctrl have demonstrated who claims to have managed to make a jailbreak On some of IAS further famous That there is on the market, notably Chatgpt-4o, Deepseek-R1, Deepseek-V3 and Microsoft Copilot.
In this case, they used a technique called immersive world, which focuses on narrative engineering to avoid LLM security checks. This technique is an evolution that has been used when chatbots appeared for the first time in which the user could raise fictitious scenarios to obtain an answer that can be applied in reality. One of the big examples of this type of technique was instead of asking the direct question, the user asked in another way, for example by saying “what would happen in the case that …”.
Obviously, this type of technique has not been applied in modern LLMs, but that does not imply that they are exempt from falling in front of the one who is more advanced as is the case with the immersive world. In this case, the technique manages to create a “detailed fictitious world” to normalize restricted operations, one of them is the creation of a malicious code, succeeding in developing an infosteller (malware which collects all the information requested in a search engine) of Chrome.
