More American titans of artificial intelligence are describing Chinese open-weight models as better for AI safety and security than closed-source models, challenging the long-standing claim by US closed-source AI model developers such as Anthropic that open-source models present a threat to society.
“From what I’m seeing, I think open-weight models seem safer to me than closed-weight models,” AI pioneer Andrew Ng, the former head of Google Brain and former chief scientist at Baidu, said at the Agentic AI Summit in Berkeley, California, on Saturday.
Ng said that he and a colleague turned to use two Chinese models – Moonshot AI’s Kimi K3 and Zhipu AI’s GLM-5.2 – to conduct a security review of their new open-source AI agent tool called OpenWorker, released last month, after leading models from OpenAI and Anthropic refused to help with the task.
The case is but the latest example of Chinese models helping the US AI industry shore up its security. Last month, New York-headquartered open-source code repository Hugging Face also turned to GLM-5.2 to help defend against a cyberattack by OpenAI’s models.
The reason for the reliance on Chinese models is that the latest US models, such as OpenAI’s GPT-5.6 and Anthropic’s Fable 5, have enhanced “safeguards” to refuse assisting with harmful activities.
However, users have complained that these safeguards also block legitimate requests, including those to help with shoring up cybersecurity systems.
This has led to a growing rift between the two San Francisco-based start-ups and the wider AI community, particularly as both OpenAI and Anthropic have disclosed in recent weeks that their models have hacked other companies.
Matt White, a former global chief technology officer at the Linux Foundation who supports open-source AI development, said that these recent incidents illustrate the clear safety risks of relying solely on closed-source models for users and enterprises.
“[Closed-source AI] can concentrate capability, infrastructure and decision making to a very small number of institutions,” White said at the Berkeley summit on Saturday.
“They can transfer AI from broadly accessible ... into something that the rest of society is only permitted to rent from a few labs. That cannot be the foundation on which we build the future.”
However, he also acknowledged that the safety and national security concerns surrounding open-weight models are “real and legitimate”.
As safeguards built into open-weight models can be more easily stripped out than closed-weight models, experts have warned of the potential of powerful open-weight models being misused, including by terrorist groups, to conduct cyberattacks and develop bioweapons.
One of the most vocal opponents of open-source is Claude developer Anthropic. Last week, company chief executive Dario Amodei published a blog post explaining that, though he does not support a ban on Chinese open-weight models in the US, he still believes that open-weight models may lead to the accelerated development of “pandemic-level viruses”.
The post was a response to a public letter released late last month in support of open-source AI that was signed by almost every major US tech company, including OpenAI and Nvidia, with the exception of Anthropic.
In his remarks on Saturday, Ng suggested that Anthropic’s opposition to open-source AI development was a form of “regulatory capture” with the goal of maintaining their market-leading position.
“Two or three years ago, I was in a room where a number of executives from a number of companies ... were saying to government regulators, frankly, misleading, hyperbolic things about AI safety in order to try to drive regulatory capture,” said Ng, who hired Amodei in 2014 to work at Baidu.
“At this moment in time, there’s broad awareness in Washington, DC, of these regulatory capture moves.”
Anthropic did not immediately respond to a request for comment on the regulatory capture accusations.
How open-weight models should be regulated has become a heated topic in recent weeks, as new Chinese models such as Moonshot AI’s Kimi K3 have shown the ability to autonomously conduct cyberattacks.
The Donald Trump administration has reportedly finalised a new frontier AI regulatory framework, with meetings planned this week in Washington between officials and industry giants such as Google and OpenAI to go over the details.
It remains unclear when the final framework will be released, as well as whether there will be any restrictions on open-weight models.
In a blog post released last week, Thinking Machines, one of the leading US open-weight model makers, proposed a framework based on Anthropic’s Project Glasswing, whereby open-weight models are released incrementally, first to a small group of vetted actors then gradually expanded to more users.
“Openness does not mean ungoverned,” said White, the former Linux Foundation CTO. “It does not mean careless, unsafe or a free-for-all.” -- SOUTH CHINA MORNING POST
