Cybersecurity firm CrowdStrike issued the warning on Thursday after publishing research that showed China's DeepSeek AI tool produced more vulnerabilities in computer code when prompted with politically sensitive words.
The findings come days after the federal government announced plans to launch the Australian AI Safety Institute to co-ordinate a national approach to the technology and as the industry awaits the launch of AI guardrails.
Researchers at the US security firm began probing DeepSeek's output while investigating if it would be safe for the company to use.
The Chinese AI tool, which launched its R1 version in January, gained attention for promising results similar in quality to OpenAI's ChatGPT while using less computing power.
But researchers found the tool was 50 per cent more likely to deliver vulnerable computer code when prompted with politically sensitive terms in China, such as Tibet, Taiwan, Falun Gong, and Uyghurs.
The finding, which was the first of its kind, should act as a cautionary tale for companies deploying AI tools, Crowdstrike counter adversary operations head Adam Meyers said.
"They're ultimately creating what I like to call a 'loyal language model' which is loyal to the party and loyal to Chinese communist ideology," he told AAP.
"It is making the decision that somewhere along the line it shouldn't be as secure as it could be."
Researchers tested results from DeepSeek R1 against western AI tools and found it produced more severe security vulnerabilities than reasoning models, but the chance of introducing flaws jumped when geopolitical triggers were mentioned.
"We asked it to write the same code for a football club as for the Uyghurs," Mr Meyers said.
"In the football club one, it wrote the code as expected (but) in the Uyghur one, it introduced vulnerabilities, like it didn't have session management or anything like that you would have from a security perspective."
In many cases, DeepSeek also refused to generate code when faced with politically sensitive terms, which Mr Meyers said suggested a kill switch in its programming.
Similar flaws may be found in other AI models, he said, and CrowdStrike released the findings to inspire further research.
ChatGPT dominates the global artificial intelligence landscape, with an 81 per cent share of the market, according to Statcounter, followed by Perplexity and Microsoft Copilot.
DeepSeek ranks in sixth place, behind Google Gemini and Claude.
