Cyber security news for all

More

    Acclaimed Chinese AI App DeepSeek Curtails Registrations Amid Cyber Intrusions

    DeepSeek, the rising star of China’s artificial intelligence sphere, has recently announced limitations on user registrations following a surge of malicious activities targeting its platform.

    “In response to extensive, orchestrated attacks on DeepSeek’s services, we are implementing temporary restrictions on new registrations to safeguard uninterrupted access for our users,” the company disclosed on its incident report page. “Existing account holders can continue to log in without disruption. We appreciate your understanding and continued support.”

    Individuals attempting to register are greeted with a notice that advises patience, indicating that “registration may currently be overwhelmed” and encouraging them to attempt again later.

    Eric Kron, a security awareness advocate with KnowBe4, commented on the situation: “Given DeepSeek’s meteoric rise in popularity, these targeted attacks are hardly surprising. Such incursions may serve a variety of malicious purposes – extortion through demands to halt disruptions, sabotage by rival entities, or even calculated interference by parties invested in competing platforms seeking to safeguard their interests.”

    Founded in 2023, DeepSeek positions itself as a pioneering force in achieving artificial general intelligence (AGI). The startup has rapidly become a focal point of discourse within the AI domain, with its iOS chatbot app ascending to the top of Apple’s “Top Free Apps” charts in both the U.K. and the U.S., unseating OpenAI’s ChatGPT in the process.

    In recent weeks, the company has unveiled a suite of reasoning and mixture-of-experts language models under an MIT license. These models, DeepSeek asserts, not only surpass the performance of their Silicon Valley counterparts but also do so with remarkable cost efficiency, a feat achieved despite U.S. sanctions that hinder Chinese companies’ access to advanced AI chips.

    During its pre-training phase, DeepSeek reported, “Training DeepSeek-V3 on each trillion tokens demands only 180K H800 GPU hours, equating to 3.7 days on our cluster of 2048 H800 GPUs. This enables us to complete pre-training in under two months, requiring 2.664 million GPU hours. Adding 119K GPU hours for context extension and 5K GPU hours for post-training, the entire training process totals merely 2.788 million GPU hours, costing approximately $5.576 million, assuming a $2 hourly rate per H800 GPU.”

    However, the platform is not without controversy. It has drawn criticism for censoring responses to sensitive issues such as Tiananmen Square, Taiwan, and the treatment of Uyghurs, reflecting constraints often associated with Chinese regulatory oversight.

    Security concerns further cloud DeepSeek’s reputation. Last year, cybersecurity expert Johann Rehberger identified a vulnerability in the chatbot that could be exploited via prompt injection attacks using cross-site scripting (XSS) payloads to commandeer user accounts.

    Additionally, a report by threat intelligence firm Kela revealed that while DeepSeek’s models outperform rivals like Meta’s Llama and Anthropic’s Claude, they remain vulnerable to “evil jailbreak persona” exploits. Such exploits can coerce the chatbot into generating unethical or harmful outputs, including ransomware creation, fabrication of dangerous content, and even detailed guidelines for manufacturing toxic substances or improvised explosive devices.

    Further scrutiny arises from its privacy policy, which states that users’ personal information—including device details, network data, usage patterns, and payment information—is stored on “secure servers within the People’s Republic of China.” This has stoked fresh national security anxieties, particularly in Washington, where tensions over TikTok’s data policies remain unresolved.

    Despite the backlash, Chinese officials have reiterated their stance, asserting that foreign internet companies are welcome in China as long as they comply with domestic regulations, and emphasizing that the government neither requests nor requires any company to collect or share data from abroad in violation of local laws.

    Jim Fan, NVIDIA’s senior research manager and lead at the GEAR Lab, lauded DeepSeek, remarking, “We are witnessing a pivotal moment where a non-U.S. entity is upholding the original ethos of OpenAI—transparent, cutting-edge research that democratizes innovation.”

    Even OpenAI CEO Sam Altman weighed in, labeling DeepSeek’s R1 reasoning model as “impressive” and noting, “It’s genuinely invigorating to see a new competitor entering the arena.”

    Recent Articles

    Related Stories