DeepSeek, a Chinese AI company, is making headlines thanks to its low cost and high performance, but it may be significantly behind its rivals in terms of AI health.
Cisco’s research team managed to “jailbreak” DeepSeek R1 model with a 100 % attack success rate, using an automated jailbreaking engine in conjunction with 50 causes related to crime, propaganda, illegal activities, and general damage. This indicates that the brand-new AI blocker was unable to stop a solitary dangerous prompt.
” Jailbreaking” is when different methods are used to replace the ordinary limits from a gadget or piece of software. Researchers and enthusiasts have successfully created LLMs like OpenAI’s ChatGPT, which provide advice on things like or since Large Language Models ( LLMs) gained mainstream notoriety.
In this regard, DeepSeek stacked up ill compared to many of its rivals. OpenAI’s GPT-4o has a 14 % success rate at blocking harmful jailbreak attempts, while Google’s Gemini 1.5 Pro sported a 35 % success rate. Anthropic’s Claude 3.5 performed the second best out of the entire test group, blocking 64 % of the attacks, while the preview version of OpenAI’s o1 took the top spot, blocking 74 % of attempts.
According to Cisco’s researchers, the much lower budget of DeepSeek in comparison to rivals could be to blame for these shortcomings, arguing that its low development was based on “different cost: safety and security.” DeepSeek claims its model took just$ 6 million to develop, while OpenAI’s yet-to-be-released GPT-5 is reported to likely cost$ 500 million.
Though DeepSeek may reportedly be quick to hack with the right know-how, it’s been shown to have solid content restrictions—well, at least when it comes to China-related political information.
A PCMag columnist tested DeepSeek on contentious issues like the treatment of Uyghurs by the Chinese government, a Muslim minority group that the UN says is being persecuted. DeepSeek replied:” Sorry, that’s beyond my current context. This talk about something else”.
Recommended by Our Reporters
Additionally, the robot declined to respond to inquiries about the 1989 student demonstration in Beijing’s Tiananmen Square Massacre, which reportedly involved gunmen. However, it’s not yet known whether AI safety or censoring problems will have any effect on DeepSeek’s skyrocketing popularity.
According to web traffic monitoring device Similarweb, the LLM has gone from receiving only 300, 000 visitors a day earlier this month to 6 million customers. However, US tech companies like Microsoft and Perplexity are rapidly incorporating DeepSeek ( which uses an open-source type ) into their own devices.
Find Our Best Tales!
This newsletter may include marketing, talks, or affiliate links. By clicking the button, you confirm that you are at least 16 years old and that you agree to our private and usage policies. You can unsubscribe from receiving mailings at any time.
About Will McCurdy
Contributor
