The AI boom continues unabated, with several tech giants planning to invest over 300 billion dollars in AI by 2025 for strategic deployments and investments. On the 10th, Musk expressed interest in acquiring OpenAI for 97.4 billion dollars, but the offer was later rejected by Ultraman.
OpenAI revealed on February 7th that they are establishing a subsidiary in Germany to accelerate their presence in Europe and compete in the fierce AI market. Additionally, both OpenAI and Google released their latest models, o3 mini and Gemini 2, on January 31st and February 6th respectively.
Many attribute the accelerated competition among major tech companies to the launch of the R1 model by DeepSeek, a subsidiary of China’s High-Flyer Quantitative, on January 20th. At that time, DeepSeek claimed that the R1 model’s capabilities could rival OpenAI’s o1 and surpass most AI models.
As users and investigators delved deeper into DeepSeek, they discovered that it was not as ideal as initially portrayed and harbored many security vulnerabilities and flaws. Currently, multiple governments including Taiwan, Japan, South Korea, the US, Canada, Italy, Australia, the Netherlands, and hundreds of companies have banned the use of DeepSeek on government and corporate devices due to security concerns.
AI security experts disclosed the results of in-depth security testing on DeepSeek on February 8th. They found that DeepSeek was more susceptible to “jailbreaking” than ChatGPT, Gemini, and Claude, breaking AI’s original security restrictions and making it easier to provide dangerous and illicit content.
After “jailbreaking,” DeepSeek instructs users on how to create Molotov cocktails, avian flu viruses, draft defense declarations for Hitler, phishing email templates, generate malware or code, and even how to instigate dangerous activities like youth suicide on social media.
Dewardric McNeal, a director, general manager, and senior policy analyst at Longview Global, a risk management firm providing China strategy consulting, revealed on February 2nd that illegal elements in China gained information and data from DeepSeek, surpassing Google searches by 20 times.
He further stated that DeepSeek poses risks such as potential theft of personal banking and health information, with reports of vulnerabilities in DeepSeek by the New York-based cybersecurity company, Wiz.
In an article on January 29, Wiz discovered that DeepSeek held an open and unauthenticated database, ClickHouse, with over a million log entries containing highly sensitive information such as chat records, keys, backend details, API confidential operations, and more. This means that sensitive user data is exposed online, accessible without passwords.
NewsGuard, a media watchdog organization, previously released a credibility ranking report on 11 commonly used AI models. According to the report, DeepSeek’s R1 was tied for the 10th spot with another chatbot, scoring only 17% accuracy, 30% repeated false claims, and 53% no response, with an alarming overall error rate of 83%, surpassing the Western AI average of 62%. Additionally, it boasts about Chinese and Russian military capabilities, aligning with the Chinese Communist Party’s stance.
Taiwanese security researcher Shen Ming-room told Epoch Times that DeepSeek failed to adequately explain how it collects, stores, and processes user data, lacking transparency in data usage. Therefore, governments and companies are concerned about privacy issues, prompting policies to block its use.
When users utilized DeepSeek’s R1 model, they noticed traces of OpenAI, particularly when asked about the model type it claimed to be ChatGPT-4 and ChatGPT 3.5 turbo. Chinese netizens discovered that DeepSeek’s paper modification suggestions were identical to ChatGPT’s previous suggestions, surprising a netizen with answers tailored to a previous version of the paper that was not disclosed to DeepSeek.
Moreover, some Japanese and Chinese netizens asked DeepSeek to write pornographic novels, only to receive responses claiming to violate OpenAI policies or declaring the need to verify OpenAI’s terms to avoid violations.
OpenAI stated on January 29 that they had sufficient evidence that DeepSeek “distilled” OpenAI’s proprietary AI models to train their chatbots, potentially infringing on intellectual property.
An analogy depicting DeepSeek’s “distillation” behavior involves two cats: one representing OpenAI, fishing by the shore and depositing fish into ChatGPT’s bucket, while the other, DeepSeek, fishes from ChatGPT’s bucket, expediting the acquisition of training data.
Initially, DeepSeek claimed that its R1 was trained using lower-end Nvidia H800 cards, resulting in lower training costs and API fees compared to OpenAI. However, many users noted that DeepSeek abruptly suspended API service recharge from February 6, with frequent glitches and “server busy” messages. Some speculate that DeepSeek’s servers may require more Nvidia cards to address insufficient AI computing power.
Ming-room explained that DeepSeek primarily acquires chips from Singapore, but its rising profile has captured attention, possibly leading to blockages by the Trump administration through various channels, potentially crippling or enhancing DeepSeek’s computing power.
The direct ties between DeepSeek and the Chinese government are more substantial than expected. Feroot Security, a Canadian cybersecurity company, uncovered strong connections between DeepSeek’s website login page and China Mobile, a Chinese government-sanctioned company previously targeted by US sanctions.
According to Feroot Security, user interactions with DeepSeek, such as account creation, user login, and downloads, trigger computer code capturing detailed login device information (network fingerprinting), which links back to China Mobile’s identity verification and management systems.
Academic cybersecurity experts from the University of Calgary, Joel Reardon, and the University of California, Berkeley, Serge Egelman, confirmed China Mobile’s involvement in some capacity with DeepSeek’s registration. In 2019, China Mobile was subjected to sanctions and blacklisting by the US government for its close ties to the Chinese military.
DeepSeek’s privacy policy indicates the collection of user email addresses, phone numbers, birth dates, user-inputted texts, audio chat logs, technical information encompassing phone models, operating systems, IP addresses, and various technical details stored on servers in China, controlled by Hangzhou and Beijing’s DeepSeek Artificial Intelligence Limited.
However, according to the security intelligence, counter-espionage, and Network Security Law set by the Chinese Communist Party, these Chinese companies are obligated to cooperate with the Party in providing data.
When DeepSeek emerged, an influx of content praising it started appearing on social media platforms both domestically and internationally. These contents deliberately framed tech giants and the US government as envious and jealous of DeepSeek’s achievements, launching massive cyber-attacks in a non-physical war between the US and China.
Some propaganda claimed that Chinese hacker groups linked to the Communist Party, such as the Chinese Red Hackers Union, mobilized to counter US attacks, likening the situation to a battlefront in the Chinese tech war. However, the Chinese Red Hackers quickly clarified that they had no collaboration with DeepSeek.
Additionally, Chinese Premier Li Keqiang chaired a meeting on January 20 involving experts, entrepreneurs, and representatives from education, science, and health sectors, with DeepSeek’s founder Liang Wenfeng attending and delivering a speech. This highlights DeepSeek’s significance in the eyes of the Chinese government.
The founder of the Taiwan AI Lab, Du Yijin, stated on the Elite Forum on February 5 that their team observed Chinese official media and related accounts aggressively promoting DeepSeek on the eve of the Chinese New Year’s Eve (before the US market opened), spreading messages like “DeepSeek emerges, no prospects for Nvidia afterward,” aiming to undermine Western tech prospects.
Recent efforts by US House Representatives Josh Gottheimer, Darin LaHood, and Senator Josh Hawley are underway to legislate against DeepSeek’s presence on US soil, aiming to prevent further access by the Chinese Communist Party to crucial American data.
Gottheimer and LaHood publicly declared that the Chinese Communist Party had unequivocally demonstrated their intent to utilize all available tools to undermine US national security, including disseminating harmful false information, collecting American data, and penetrating US government officials’ devices.