US AI leaders turn to Chinese open-weight models, challenging closed-source safety claims

American artificial intelligence heavyweights are viewing Chinese open-weight models as a safer alternative to closed-source counterparts, challenging the notion that open-source models pose a threat to society, according to reports. Andrew Ng, a prominent AI expert and former Google Brain head, stated that Chinese models, such as Moonshot AI's Kimi K3 and Zhipu AI's GLM-5.2, appear safer than closed-weight models.

Ng and a colleague utilized these Chinese models to conduct a security review of their open-source AI agent tool, OpenWorker, after leading US models refused assistance. Similar reliance on Chinese models has been observed across the industry, with Hugging Face, an open-source code repository, also turning to GLM-5.2 for cybersecurity defense against OpenAI's models.

The preference for Chinese models arises due to the enhanced "safeguards" in recent US models, which often block legitimate requests, including those aimed at enhancing cybersecurity. This reliance has caused a rift between US-based AI start-ups, such as OpenAI and Anthropic, and the broader AI community. Matt White, a former global CTO at the Linux Foundation, emphasized the safety risks associated with closed-source models, stating that their concentration of capability, infrastructure, and decision-making to a few institutions could be problematic.

However, White acknowledged that open-weight models' safety and national security concerns are "real and legitimate." Critics, like Anthropic CEO Dario Amodei, argue that open-source models may hasten the development of "pandemic-level viruses." Ng suggested that Anthropic's opposition to open-source AI may be a strategy to maintain their market-leading position, describing it as "regulatory capture."

The US government is reportedly finalizing a new frontier AI regulatory framework, with potential restrictions on open-weight models. A leading US open-weight model maker, Thinking Machines, proposed a framework for incremental open-weight model releases to address safety concerns, emphasizing that openness does not equate to ungoverned or unsafe practices.