Loading Logo

Monsters in the deep: AI disruptor DeepSeek is an impressive challenger with a big privacy problem

March 2025
 by Sarah Keeling

Monsters in the deep: AI disruptor DeepSeek is an impressive challenger with a big privacy problem

March 2025
 By Sarah Keeling

DeepSeek, the Chinese LLM that overtook ChatGPT as the most downloaded free app at the beginning of 2025, has earned its reputation as a disruptor. The open-source model has threatened the dominance of leading AI companies by offering an alternative where datasets, algorithms and models aren’t shrouded in secrecy. The cost and energy efficiency of the chatbot has undermined the assumption that generative AI depends on a high supply of GPUs. At a time when big-tech and the environment seem to be mutually exclusive, it offers a possible route to greener AI use. It can even tell you how many “r”s there are in “strawberry”. However, the impressive capability of the model is not without its flaws, especially with regards to privacy on both the DeepSeek app and website.

Before DeepSeek came to international attention, most of the familiar names in or behind AI were US-based: OpenAI, Anthropic, Meta, Alphabet, Perplexity, X. DeepSeek, on the other hand, is firmly Chinese. Its founder, Liang Wenfeng, co-founded the Hangzhou-based hedge fund High-Flyer, which wholly owns and funds DeepSeek. Founded in 2003 as an offshoot AI lab separate from the business’s main focus on AI-driven trading algorithms, every model DeepSeek has produced has been concurrent with US sanctions on chip exports, reportedly using a stockpile of Nvidia GPUs acquired in 2021, raising doubts as to whether these trade restrictions have produced the desired effect. The chatbot also has a few habits which give away the influence of China’s government on the model; users asking questions about Tiananmen Square and the Umbrella Revolution have been met with a blank “Sorry, that’s beyond my scope. Let’s talk about something else”. While users have increasingly been discovering workarounds for these obstructions, there is less wiggle room under China’s National Intelligence Law. DeepSeek’s privacy policy confirms that it stores data on servers located in the PRC. Under China’s 2017 cybersecurity laws, DeepSeek could be compelled to share data with the government upon request. Other Chinese chatbots such as search engine Baidu’s Ernie Bot and Moonshot’s Kimi AI have also been criticised on similar grounds; in the wake of DeepSeek’s January 2025 surge in popularity both have released free models.  Conversely, Mistral’s Le Chat from France, launched in February 2025, is drawing interest as a competitively priced, yet GDPR-compliant option.

Whether DeepSeek sends your prompts and conversations to the state aside, the site has also been targeted by hostile actors worldwide. From 27 January to the first week of February 2025, DeepSeek also reported numerous “large-scale” DDOS cyberattacks on both the API and web chat service, causing the site to temporarily pause new user registrations. In addition to these concerns, DeepSeek has been accused of lax data security practices. On 29 January, Wiz, a cybersecurity firm based in New York, highlighted an exposed database of over one million records. The ClickHouse database was unsecured, unauthenticated, and publicly accessible; “within minutes” Wiz was able to gain access not only to chat and prompt history, but passwords, API authentication tokens, local files and proprietary information.

After Wiz notified DeepSeek of this exposure, the site reportedly fixed the issue within an hour. To those affected, this should not be taken as a cue to relax, as it is highly likely that other entities discovered the leak before it was officially communicated to DeepSeek. The researchers at Wiz were concerned by how basic the oversight was; Wired reported that the level of exposure is the sort of thing that Wiz typically finds after hours of searching through neglected services, emphasising that DeepSeek has failed to meet the “bare minimum” in defending against vulnerabilities.

The unholy trinity of fears over practice, external attacks, and the geopolitical tensions of China’s potential to access data mean that users should be extremely cautious about using both the API and web chat services, especially for proprietary information. Concerns over the chatbot’s data collection practices have already resulted in international action. At the end of January 2025, Italy’s data protection authority ordered a countrywide block, and Australia, Taiwan and the state of New York have banned the use of the app on government devices. The European Data Protection Board also met on 11 February to discuss DeepSeek, and South Korea has banned new downloads of the app, although existing users can still use the chatbot and the website remains accessible.

While the model may be operating on a very impressive level for the computing power it uses, at present it is hard to be confident in DeepSeek’s data. For those who have already used DeepSeek, crucially before the database leak was addressed, increased vigilance is recommended against known threats such as phishing. It may also be worth updating credentials. Regardless of the origin of a promising model, caution should always be exercised when trusting them with your sensitive data.

Join our newsletter and get access to all the latest information and news:

Privacy Policy.
Revoke consent.

© Digitalis Media Ltd. Privacy Policy.