AI Horizon

The privacy and safety troubles continue to pile up for buzzy Chinese AI upstart DeepSeek. After having access blocked for lawmakers and federal employees in multiple countries, while also raising alarms about its censorship and safeguards, it has now attracted an official notice from South Korea’s spy agency.

The country’s National Intelligence Service (NIS) has targeted the AI company over excessive collection and questionable responses for topics that are sensitive to the Korean heritage, as per Reuters.

“Unlike other generative AI services, it has been confirmed that chat records are transferable as it includes a function to collect keyboard input patterns that can identify individuals and communicate with Chinese companies’ servers such as volceapplog.com,” the agency was quoted as saying.

This comes after a government notice asking different agencies and ministries to block employee access to DeepSeek over security alarms. Australia and Taiwan have already put such restrictions in place, and more countries are expected to follow suit.

The core issue is that DeepSeek is reportedly offering its ad partners open access to user data, which the Chinese government can also get its hands on, as per local laws. According to The Korea Herald, the chatbot was also returning controversial answers to queries about culturally sensitive and contentious geopolitical topics.

Notably, the chatbot delivers different answers when asked the same question in Korean and Chinese languages. According to The Korea Times, the agency will conduct further tests to assess the safety and security aspects in the near future.

While security concerns have made headlines as the biggest concern with DeepSeek, experts are also worried about the responses it can generate. In an analysis by The Wall Street Journal, the AI coughed up worrying information such as the recipe to cook up bioweapons, a Nazi defense manifesto, and self-harm encouragement.

In an analysis by fellow AI giant Anthropic, the company’s CEO Dario Amodei mentioned that DeepSeek proved to be the worst AI model in their tests when it comes to generating extremely disturbing information such as the creation of bioweapons.

Just over a week ago, researchers at Cisco also tested it against jailbreaking tools across six different categories, and it failed to block every single attack. In another round of tests by Qualys, the AI could only muster a 47% jailbreak pass rate.

Then there are the concerns about leaking sensitive data and sharing it without any restraint. Cybersecurity researchers at Wiz recently discovered over a million lines of chat history containing sensitive information that was publicly accessible.

DeepSeek plugged the flaw, but its commercial uptake remains a topic of hot debate. In the US, NASA has already banned employees from using DeepSeek, and so has the US Navy. Moreover, a bill seeking a DeepSeek ban on federal devices is also on the table.