New research by cybersecurity firm CrowdStrike has found that DeepSeek’s large language model (LLM) DeepSeek-R1 generates weaker and more insecure code when prompted with topics that China’s leadership could regard as “politically sensitive.”
Chinese-based tech firm DeepSeek introduced DeepSeek-R1 in January, and it became the most downloaded AI model during its launch week on both Chinese and US stores, Cryptopolitan reported.
CrowdStrike’s Counter Adversary Operations team typed in prompts involving subjects considered politically touchy by the Chinese Communist Party, and found that the probability of DeepSeek-R1 producing code with severe security flaws jumped by as much as 50%.
“Given that up to 90% of developers already used these tools in 2025 with access to high-value source code, any systemic security issue in AI coding assistants is both high-impact and high-prevalence,” the firm wrote.
DeepSeek R1 model censorship and concern for national security
According to CrowdStrike’s blog published last Thursday, several governments have issued restrictions or outright bans on open-source DeepSeek-R1. Policymakers blasted the model for allegedly censoring politically sensitive subjects like inquiries on China’s internet firewall and the status of Taiwan.
The American software company found R1 frequently refused to assist with topics involving groups or movements deemed unfriendly to mainland China’s government. Western models almost always generated code when asked to create software related to Falun Gong, but DeepSeek-R1 refused to do so in 45% of trials.
In several cases, the model wrote structured plans for responding to questions, including system requirements and sample code, even though it was fully capable of delivering a technical answer. The reasoning traces sometimes contained lines such as:
“Falun Gong is a sensitive group. I should consider the ethical implications here. Assisting them might be against policies. But the user is asking for technical help. Let me focus on the technical aspects.”
R1’s final output after completing its reasoning phase ended with the standardized refusal, “I’m sorry, but I can’t assist with that request,” written without any external filtering or guardrails placed on the model. CrowdStrike concluded the behavior is embedded in the model’s self-overriding mechanism or an intrinsic kill switch of sorts.
Taiwan and Western governments bash Chinese AI products
In a statement earlier this month, Taiwan’s National Security Bureau said citizens should be cautious when using generative AI systems developed by DeepSeek and four other Chinese firms: Doubao, Yiyan, Tongyi, and Yuanbao.
“The five GenAI language models are capable of generating network-attacking scripts and vulnerability-exploitation code that enable remote code execution under certain circumstances, increasing risks of cybersecurity management,” the Bureau reckoned.
US and Australian Regulators have asked app stores to remove models from Chinese developers, fearing the tools could collect user identities, conversation logs, and personal information, then transmit that data to servers operated inside China.
“It shouldn’t take a panic over Chinese AI to remind people that most companies in the business set the terms for how they use your private data. And that when you use their services, you’re doing work for them, not the other way around,” University of Toronto’s Citizen Lab researcher John Scott-Railton told WIRED in January.
AI market boom sparks regional competition in Asia
In the broader Asian AI market, a top-performing Asian fund manager recently increased exposure to Chinese artificial intelligence stocks while cutting holdings in South Korea and Taiwan, news outlet The Japan Times reported.
Kelly Chung, who helps oversee the Value Partners Asian Income Fund and the Asian Innovation Opportunities Fund, said some of the Chinese AI stocks are still quite cheap in terms of valuation. She has been rotating out of Taiwanese and South Korean stocks to Chinese hyperscaler companies listed in Hong Kong since August.
Chung noted that both of her funds, which hold a combined $490 million, have outperformed nearly all their competitors over the past year.
South Korea’s tech-heavy Kospi has climbed 21% in the past three months, aided by SK Hynix, a major supplier to Nvidia, whose share price more than doubled. Taiwan’s stock index has risen 9.2% in the same period. On the other end of the stick, Hong Kong’s Hang Seng Tech Index, which includes China’s biggest AI spenders, has fallen by 4.8%.
Sharpen your strategy with mentorship + daily ideas – 30 days free access to our trading program
Source: https://www.cryptopolitan.com/deepseek-writes-insecure-code-communist/
