Arcee: Chinese open AI models pose no security threat

Experts at the US-based Arcee laboratory consider concerns regarding the security of open-source AI models developed in China to be unfounded. Recently, models presented by Chinese companies such as Moonshot AI and Alibaba have gained popularity in the global market due to their efficiency and affordability, sparking various discussions among Western tech giants. This is reported by Techcrunch.com reports .

At a time when the US administration is considering banning products from Chinese developers, creators of closed models like OpenAI and Anthropic are concerned about competition. Specifically, models like Kimi K3 and Qwen stand out for providing significantly cheaper data processing services. This could have a noticeable impact on the profit margins of large corporations.

According to Lucas Atkins, CTO of Arcee, Chinese open-weight models are no more dangerous than any other open-source software. Atkins emphasizes that enterprises using these models do not need to fear attacks from Chinese hackers, as the operating principle of neural networks differs from traditional software.

Technical security and control capabilities

Many perceive these models as code written with malicious intent, but the AI training process is entirely different. Arcee experts explain that if a company downloads a model and runs it on its own private servers (data centers), the party that created the model (e.g., Alibaba) will not have access to that environment or the ability to steal data.

The operating code of open-weight models hosted on platforms like Hugging Face is transparent and verifiable. Although the methods and datasets used to train the models remain proprietary, experts can fully control the part running on the server. This opens the door to implementing security measures independently.

Large organizations should subject any model to security testing before deployment, checking for aspects such as bias and hallucinations. Furthermore, the reliability of models can be further enhanced through post-training to suit specific needs.

Interestingly, Arcee actually offers US-made domestic products as alternatives to Chinese models. Nevertheless, they avoid making unfounded accusations against competitors, demonstrating their support for technological openness. This approach serves to foster healthy competition and innovation in the global AI market.