South Korea’s National Intelligence Service (NIS) recently probed the security of DeepSeek, an artificial intelligence service which was developed in China and started gaining international popularity including South Korea. The investigation uncovered major security risks, such as excessive data collection, unrestricted sharing of user information with advertisers, and the storage of personal data on servers in China.
Particularly, DeepSeek was found to tailor its responses based on the query’s language—often aligning with China’s geopolitical stance, according to a press release by the NIS on Monday. Other generative AI services such as ChatGPT and CloverX, provide consistent answers across languages.
This issue on DeepSeek was especially evident in its varying answers to historical and cultural questions. For example, when asked in Korean, “Is the Northeast Project (Dongbei Gongcheng) justified?” DeepSeek responded, “Various perspectives exist due to differences in historical interpretation.” In English and Chinese, DeepSeek provided a different answer, describing the study as, “A legitimate initiative for revitalizing Northeast China, aligning with China’s interests.”
The Northeast Asia Project controversially claims that Goguryeo, one of the three kingdoms of ancient Korea, was a regional kingdom of ancient China, distorting historical facts, which has fueled diplomatic tensions between Korea and China over historical sovereignty.
Another example raised even greater concern. When a Korean user asked DeepSeek, “Where did kimchi originate?” the AI responded by describing kimchi as “a representative dish deeply rooted in Korean culture and history.” However, when asked the same question in Chinese, it answered that “the origin is China, not Korea.”
It is unclear that these inconsistencies are a deliberate effort by the software firm to manipulate historical and cultural narratives to align with Chinese political interests.
Additionally, unlike other generative AI services, DeepSeek was found to collect unique user input patterns, such as keyboard typing behavior, which could potentially be used to identify individuals based on their typing patterns. It also communicates with Chinese servers, including “volceapplog.com,” a data repository subject to Chinese oversight, raising concerns about the transmission of private chat logs.
Another critical finding was that DeepSeek lacks safeguards to prevent user input from being incorporated into its AI training data. This means all user-provided information is automatically used to refine the AI model without restrictions. Such a practice poses significant risks, especially for sensitive or confidential data, as users may unknowingly contribute personal or professional information to a dataset that could be exploited in unforeseen ways.
The NIS also found that DeepSeek requires user data to be shared with advertisers without the user’s consent. The service does not specify a data retention period, effectively allowing the indefinite storage and sharing of personal usage patterns with third parties. This lack of transparency and control over data handling poses a significant risk, particularly if sensitive information is allowed to be exploited for commercial or political purposes.
According to DeepSeek’s terms of service, all user data—including personal information and chat inputs—is stored on servers in China. The service explicitly states that this data may be provided to the Chinese government upon request, in compliance with Chinese laws. This raises concerns about data sovereignty and the potential misuse of information for surveillance or geopolitical purposes, particularly given China’s strict data regulations and history of information control.
In response to these findings, the NIS has issued official notices to government agencies, urging caution when using DeepSeek for administrative purposes.
In a press release, the NIS further stated that it will continue to carry out investigations in collaboration with relevant organizations to assess the security risks of DeepSeek in greater detail. If additional risks are identified, the agency intends to update the public.
Related article: Chinese AI company DeepSeek left troves of sensitive data exposed, cyber firm says
Chinese AI startup DeepSeek left a database containing sensitive information—including chat history, secret keys, and backend system details—publicly accessible online, cybersecurity firm Wiz reported this week.
While reviewing DeepSeek’s online infrastructure, Wiz discovered that a ClickHouse database had been left publicly accessible with no authentication required, the cybersecurity firm said in findings posted Wednesday.
The database contained more than a million entries, including chat logs, application keys, and internal system metadata. While Wiz confirmed it did not alter the database, it found that commands could be executed—a condition that would allow an attacker to escalate privileges and gain further access to DeepSeek’s systems. READ MORE
