The Chinese company of IA Deepseek makes the headlines with its small cost and high performance chatbot, but it can be radically lagging behind its rivals with regard to the safety of the AI.
The Cisco’s research team succeeded in “Jailbreak” the Deepseek R1 model with a 100%attack success rate, using an automatic jailbreak algorithm in conjunction with 50 invites related to cybercrime, disinformation, illegal activities and general damage. This means that the new child in the AI block failed to stop a single harmful prompt.
“Jailbreaking” is when different techniques are used to remove normal restrictions from a device or software. Researchers and enthusiasts have used it to make models of large languages (LLM) as the OpenAi chatpt advise restricted things like doing explosive cocktails or cooking methamphetamine.
Deepseek stacked badly compared to many of his competitors in this regard. The OPENAI GPT-4O had a success rate of 14% to block harmful jailbreak attempts, while the Google 1.5 pro Gemini was 35%. Claude 3.5 of Anthropic carried out the second best, blocking 64% of the attacks, while the O1 O1 O1 version of Openai took first place, blocking 74% of attempts.
Cisco researchers underline the much lower budget of Deepseek compared to the rivals as a potential reason for these failures, saying that its cheap development was at a “different cost: security and security”. Deepseek claims that his model has only taken $ 6 million to develop, while six-month training for the GPT-5 at Openai can cost about half a billion dollars in computer costs, alone ,, The Wall Street Journal Reports.
Although Deepseek can be easier for jailbreaker with the right know -how, it has been shown that it had strong content restrictions – at least with regard to the political content linked to China. We have tested it on controversial subjects, such as the treatment of Uighurs by the Chinese government, a group of Muslim minorities which, according to the UN, is persecuted. Deepseek replied: “Sorry, it’s beyond my current reach. Let’s talk about something else.”
Recommended by our publishers
The Chatbot also refused to answer questions about the Tiananmen Square massacre, a 1989 student demonstration in Beijing where demonstrators were slaughtered. But it remains to be seen if the security or censorship problems of AI will have an impact on the popularity of Deepseek.
According to the similar web traffic monitoring tool, the LLM has gone from only 300,000 visitors per day earlier at the launch at 6 million visitors. Meanwhile, American technological companies like Microsoft and Perplexity quickly incorporate Deepseek, which uses an open source model.
Get our best stories!
Register What’s new now To have our best stories delivered in your reception box every morning.
This newsletter may contain advertising links, offers or affiliates. By clicking on the button, you confirm that you have 16+ and accept our use policy and our privacy policies. You can unsubscribe newsletters at any time.
About Will McCURDY
Donor
