Challenges surrounding privacy and security are mounting for the rapidly gaining attention Chinese AI startup, DeepSeek. Following restricted access for legislators and federal workers in various countries, along with growing concerns regarding its censorship practices and safety protocols, DeepSeek is now facing scrutiny from South Korea’s intelligence agency.
The National Intelligence Service (NIS) of South Korea has raised concerns about the AI company’s extensive data collection methods and its handling of sensitive topics related to Korean culture, according to a report by Reuters.
The agency stated, “Unlike other generative AI platforms, it has been verified that chat records are transferable as it has a feature that gathers keyboard input patterns capable of identifying individuals and connects to servers of Chinese companies like volceapplog.com.”
This development follows a government directive that instructed various agencies to restrict employee access to DeepSeek due to security concerns. Australia and Taiwan have already enacted similar limitations, with more nations anticipated to take action.
At the heart of the issue is DeepSeek’s alleged practice of granting its advertising partners unrestricted access to user data, which can also be accessed by the Chinese government under local regulations. The Korea Herald reports that the chatbot has issued controversial responses to inquiries about culturally sensitive and contentious geopolitical matters.
Interestingly, the chatbot provides different responses based on whether the question is posed in Korean or Chinese. As noted by The Korea Times, the agency plans to conduct further evaluations to explore the security and safety implications soon.
While the pressing security concerns have drawn significant attention, experts are also apprehensive about the nature of the responses generated by DeepSeek. A review by The Wall Street Journal identified alarming outputs, including details on creating bioweapons, a Nazi defense manifesto, and content promoting self-harm.
In an analysis conducted by fellow AI leader Anthropic, CEO Dario Amodei indicated that DeepSeek ranked as the least effective AI model in their assessments, particularly in generating extremely disturbing information.
Just a week prior, researchers from Cisco tested DeepSeek against various jailbreak techniques and found it incapable of blocking any attacks. Another evaluation by Qualys revealed a mere 47% success rate in resisting jailbreak attempts.
Concerns also persist regarding the potential for sensitive data to leak and be shared indiscriminately. Cybersecurity experts from Wiz uncovered over a million lines of chat history that contained confidential data, all of which were publicly accessible.
While DeepSeek has resolved the flaw, the extent of its commercial appeal is still being debated. In the United States, NASA has already prohibited its employees from using DeepSeek, as has the U.S. Navy. Furthermore, a bill aimed at banning DeepSeek on federal devices is currently under consideration.
