Arcee, a US open source AI lab, says Chinese models are not inherently dangerous

Published 22 July 2026 · tech

As Chinese open-weight AI models grow in capability and popularity, arguments about what should be done about them have once again reached a fever pitch

As Chinese open-weight AI models grow in capability and popularity, arguments about what should be done about them have once again reached a fever pitch. There’s talk that the Trump administration might try to ban them (though it hasn’t yet acted on the idea). Meanwhile, proprietary model makers, particularly OpenAI and Anthropic, appear increasingly concerned about them. Open-weight models such as Moonshot AI’s Kimi K3 or Alibaba’s Qwen offer inference at a fraction of the token cost of closed source models from these large U.S. labs. The fear is that they also pose some sort of threat. Certainly they threaten the profit margins of the large proprietary AI labs. But should enterprises running these models in their own data centers succumb to the fear that they could be a vector for Chinese hackers? No, says Lucas Atkins, the CTO of Arcee, which is building open models to give U.S. companies a homegrown alternative to Chinese models. If any startup would benefit from a ban on Chinese models, Arcee would. But Atkins says China’s open models are no more dangerous than any other open source software a company may use.

In fact, he says, they even offer benefits even to his own company. “A lot of people view this as similar to a Chinese software program. Like, it was coded with these x, y, z intentions” that a bad actor could simply command, he said. “That is fundamentally not how these models are trained. There is really not any way for an Arcee, or an Alibaba, to make a model, have someone run it in their own environment and for us have any access to it whatsoever,” he explained. While most of these models are what’s known as “open weight” and are not really fully open source software, the source code (the part that will actually run on servers), if it is downloaded from open source sites like Hugging Face, is similarly largely visible and reviewable. (What isn’t available is the methods and data used to train the models.) Large organizations should put any model core through their security testing and inspection processes, and they will also often post-train the models for their specific uses and can examine areas like bias, toxicity, hallucinations, and sensitivity to certain topics.

So they work with, optimize, and understand the models before people start sending them prompts. Could a model that is used for coding somehow throw malicious backdoors into the code it writes? Again, while that’s theoretically possible, it would require acrobatic feats to accomplish. “There’s no reason that a sophisticated enough actor couldn’t train a model to be a completely amazing coding model in every circumstance, but when presented with a certain type of code base … some hidden training would kick in,” Atkins, who spends his days training models, postulated. But he adds: “I don’t know how you would do this.” Because large language models are by nature creative, the odds are slim of getting a contemporary model to spit out malware in response to a preplanned perfect storm of context and prompt. Even slimmer are the chances that any enterprise would then use that code. Could it happen in the future? That’s anyone’s guess. But enterprises are also building their AI apps to be model-agnostic and to use multiple models.