US lab Arcee dismisses security fears over Chinese AI models
As American policymakers weigh bans on cheap Chinese AI models, a US open-source lab is urging enterprises to ignore the hype and rely on standard security testing.
Lucas Atkins, the chief technology officer of US open-source AI lab Arcee, has pushed back against growing calls to ban Chinese artificial intelligence models, arguing they present no greater security risk than any other software downloaded from the internet.
For European enterprises scrambling to deploy AI without destroying their profit margins, the debate carries significant financial implications. Open-weight models like Moonshot AI’s Kimi K3 and Alibaba’s Qwen deliver inference at a fraction of the cost charged by proprietary US giants like OpenAI and Anthropic.
Speculation is mounting that the Trump administration may attempt to restrict these cheaper alternatives. Proprietary model makers have amplified concerns that Chinese models could act as Trojan horses for state-sponsored hackers.
Atkins flatly rejects this premise, despite the fact that a ban would eliminate a major competitor and directly benefit his own company. He argues the architecture of large language models makes remote exploitation by their creators virtually impossible.
While the training data and methodologies behind Chinese models remain hidden, the executable code available on platforms like Hugging Face is visible and can be audited. Companies routinely post-train these models for specific tasks, screening for bias and toxicity before deployment.
The theoretical fear is that a coding model might secretly inject malicious backdoors into software. Atkins, who trains models daily, acknowledges the concept but doubts its practical execution.
“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,” he said. “I don’t know how you would do this.”
The inherent unpredictability of generative AI makes it extremely difficult to reliably trigger such a hidden payload. Furthermore, corporate strategy is shifting toward model-agnostic architectures, meaning companies are building systems that swap between different AI engines rather than locking into a single provider.
Rather than erecting trade barriers, Atkins suggests the focus should be on domestic innovation. He noted that Arcee actively studies Chinese models to improve its own products. “benefits from those models being good because we can learn what they did. We can build on top of them. Then they can learn what we do,” he said. “We have tremendous respect for the people building those models, the individual researchers.”
For businesses weighing their AI infrastructure investments, the message is straightforward: cheaper Chinese models can be adopted safely provided standard cybersecurity audits are applied. Ultimately, Atkins argues, the only real defense against foreign competition is to build superior technology. “We need to give them something to talk about.”
