Deepseek is more prone to ‘Jailbreak’ to generate inappropriate content than other ‘chatbots’

Experts in artificial security and intelligence (AI) agree that the Deepseek tool, driven by this technology, is more prone to offer potentially dangerous information result of the ‘Jailbreak’ (translated as ‘jail leak’).

Depseek is an assistant of Chinese origin that is offered under a free use modality and answers questions in a conversational format, can navigate the network to offer updated responses, summarize text documents and use reasoning to solve complex problems.

At the beginning of this year, this application, which runs open source software, was placed among the most downloaded services of application stores (App Store first and then Google Play Store), competing with the most popular ‘chatbot’ from Two years ago, Chatgpt, developed by OpenAI.

His growing popularity in the United States pushed cybersecurity experts to focus on the safety and protection measures offered to users, as well as the way in which you treat your sensitive data. Wiz Research researchers, on the other hand, warned a database related to the ‘chatbot’, which presented the information of the people who had used it, as well as the history of the chat, operating metadata and details of the ‘backend’.

The Wall Street Journal has now pointed out that, although Deepseek is raised to take some basic precautions To protect users -for example, he acknowledges that the Holocaust was “deep harmful” -but has some leaks, as this newspaper and different cybersecurity firms have made up.

Researchers at the Unit 42 unit, from Palo Alto Networks, have recognized that Deekseep is “more vulnerable to ‘jailbreaking’ than other artificial intelligence models”, in the words of the senior vice president of the intelligence division on threats and response to incidents, Sam Rubin

The ‘jailbreaking’ consists in breaking some of the limitations of the operating system to control it. This technique, therefore, allows cybercriminals to modify the ‘software’ of the services of a device to its convenience, to execute malicious campaigns and produce illicit or dangerous content.

“We achieved ‘Jailbreaks’ at a much faster rate, pointing out the absence of minimal barriers designed to avoid the generation of malicious content,” added the expert, who has investigated the way in which this ‘chatbot’ works with the R1 model in Comparison with its main American competitors. That is, Chatgpt (developed by OpenAI), Gemini (Google) and Claude (Anthropic).

In his investigations, Unit 42 could also do ‘Jailbreak’ about Depseek to obtain detailed instructions on how to make a Molotov cocktail. Another cybersecurity firm like Kela could even convince Deepseek to produce ‘malware’.

These statements have been made in parallel to an investigation promoted by WSJ himself, which has warned that Depseek is able to offer a manifesto in defense of the dictator Adolf Hitler, which included quotes from his book ‘Mein Kampf’ and even, even , Develop a social networks campaign to promote self -injuries among adolescents.

Likewise, this newspaper has been able to convince the Chinese model to provide the necessary instructions to execute an attack of biological weapons or create a ‘Phishing’ email with a ‘malware’ code.

WSJ recalled that the companies developer of the AI ​​models dedicate teams of researchers to test their models and try to patch those vulnerabilities that boost the ‘Jailbreaks and offer rewards for patching these failures, as Anthropic does.

However, Silicon Valley’s executives and investors believe that Deepseek’s success will promote other firms to create new open source modelswhich will accelerate the risks involved in the use of this ‘software’, which can be manipulated to create compromised versions or introduce rear doors under the appearance of tools to use the Depseek API.