Updates on Chinese AI: Kimi-K3, Xi at WAIC, and 4 Months to Mythos

Alex Colville

Jul 22, 2026

Hello, and welcome to AI Under Heaven! It’s a new newsletter at ASPI that uses primary sources to track the Chinese AI ecosystem, with a particular focus on how the technology is being used to pursue the nation’s domestic and international strategies.

I’m using an old-meets-new theme for this. While working for China Media Project, I saw a flood of Chinese state media content trying to merge traditional Chinese culture with AI, a merger of imperial and modern China that Xi Jinping has encouraged. The link between this cutting-edge technology and man’s age-old drive for global dominance is increasingly hard to ignore, so I thought I might put my own spin on the idea.

I’m still feeling out the structure, but my aim is to take any developments around Chinese AI that I think are important, and analyse them for you here. These developments will be both from within the country and overseas, as well as any interesting technical news. There may also be one-off specials on stories I think are important. Feel free to reach out with ideas for things you’d like to see.

A little about me: I’m an Analyst at the Australian Strategic Policy Institute (ASPI), focusing on China’s AI ecosystem. For the past two years I covered this beat for David Bandurski at China Media Project, where I wrote a newsletter called China Chatbot. I have a background in coding from a scholarship with the Lede Program for Data Journalism at Columbia University, and was based in Beijing from 2019 to 2022, where I worked as a staff writer for The World of Chinese. I have also written on Chinese affairs for The Financial Times, The Economist and The Wire China.

It won’t surprise you that this first issue will mostly talk about the World AI Conference (WAIC) that ran July 17-20 in Shanghai. As I wrote in ASPI’s Cyber and Tech Digest last week, this is a shopfront —China’s chance to present their AI products and governance solutions in a way international audiences can understand.

In this issue:

What did Xi mean when he endorsed “open-source” AI at the WAIC?

Who signed up to the World AI Cooperation Organization?

The UK’s AISI discovers that all you need to do to get around DeepSeek’s safety guardrails on cybersecurity…is just to ask again.

But First…

Events at the World AI Conference last week sparked further debate about the future of China’s open-source strategy. Is it opening up, or closing off? Both have a place in the country’s national interest, so I don’t think it’s either-or.

Chinese AI labs are known for releasing their models to the general public for free, which has enabled their rapid adoption by other AI developers around the globe. But estimates suggesting a Chinese lab could be as little as four months away from creating a product like Anthropic’s Mythos — a model too powerful to release to the general public — has fueled speculation that China’s open-source era might be drawing to a close.

Chinese President Xi Jinping delivers the keynote speech at the World AI Conference in Shanghai in July 2026

Chinese leader Xi Jinping seemed to answer that question during his speech at WAIC, where he endorsed “open-source and openness” as the future for international AI development and governance. Newspapers like the Wall Street Journal interpreted this to mean that Xi was endorsing open-source AI models. As a piece of signalling, it was unmistakable: openness is the way forward. On July 19, Alibaba announced that its latest model, Qwen-3.8, would be open-source. The previous iteration, Qwen-3.7, had been closed-source.

However, it’s important to remember that a Western concept can be imbued with unique political overtones in Chinese. In his speech, Xi did not mention “open-source” per se, but instead a four-character political slogan, “open source and openness” (开源开放).

The slogan itself first appears along with the first official Chinese policy document on AI from 2017, which sketches out China’s international policy for the technology, and seems to be a broad term referring to sharing any research or technology that could contribute to China’s AI development. It is still interpreted this way by state media: Xinhua used it in March to talk about open-sourcing AI chip architecture.

So what Xi promised in his WAIC speech actually has more to do with the idea of open-source, rather than specifically guaranteeing frontier AI models will always be open-source. They may be part of it, but I would argue the term is broad and vague enough to allow for frontier AI models to be taken off the table should the need arise. If Reuters is correct that the leadership has been discussing ways to restrict international access to some Chinese AI products, we may see an ecosystem that has a place for both open and closed-source AI in the future.

For more details, you can read all about it here, in ASPI’s The Strategist blog.

And on that note, on with the show. Enjoy!

Alex Colville (Analyst, ASPI)

Share

All Eyes on the World AI Conference

The World AI Conference (WAIC), a showcase of Chinese AI products and governance methods for the international community, dominated AI-related government actions and headlines in Chinese state media last week. Each government ministry in charge of AI development released a policy document staking a claim to the future international development and governance of AI.

The National Development and Reform Commission released an Action Plan for developing AI capacity across the world.

The Ministry of Industry and Information Technology released their own Action Plan for building an “ethical” AI governance framework for the international community, saying they would work with international organizations to implement it.

The Cyberspace Administration of China released a “Global Cooperation Initiative” for building trustworthy AI agents, including “establishing a security governance framework covering the entire lifecycle of intelligent agents.”

The conference received heavy coverage from key state media outlets (which inevitably included AI-generated content), who highlighted that China was prioritizing openness, safety, and inclusivity in AI’s development.

“A single string cannot produce a sound, and a single tree cannot make a forest,” wrote the “Harmony” column of People’s Daily, an editorial dedicated to international affairs. “The development of artificial intelligence should not be a solo performance by any one country, but rather a symphony of global cooperation.”

So What? Each major government department responsible for the regulation and development of AI in China has just declared it will become more involved with the international community going forward. Expect AI to be one of the best global networking opportunities these departments have ever had.

“CCP” Has Entered the Chat

On July 15, the Chinese government tightened its control over personalized AI chatbots. The “Interim Measures for the Management of Anthropomorphized AI Interaction Services” (人工智能拟人化互动服务管理暂行办法) were issued in April by the Cyberspace Administration of China (CAC) together with the National Development and Reform Commission and the Ministry of Public Security, but only became enforceable last week.

In the past, AI labs allowed users to create AI chatbots with particular personas based on their AI models, acting as a companion for them. Now any platform hosting chatbots with over 100,000 monthly active users must submit a security assessment report of their chatbots to their local CAC.

Ahead of the deadline, ByteDance, Alibaba, and Tencent all shut down this feature, essentially ceding control of chatbot creation to the platforms. This was a sizeable cull, given ByteDance’s platform alone hosted over 8 million tailored chatbots by May 2024. The Shanghai CAC has already shut down 14,000 such chatbots.

Chinese media celebrated how the measures ban the more harmful elements of AI chatbots:

No virtual partners designed to cultivate dependency.

No posing as a doctor or financial adviser delivering serious advice without credentials.

No emotional manipulation targeting underage users.

No romanticizing self-injury or suicide.

Each is a genuine issue for AI chatbots across the world, with OpenAI embroiled in an ongoing lawsuit in San Francisco, ChatGPT accused of aiding a 16-year old boy to commit suicide. However the Interim Measures also require that personalized chatbots must not “generate content that endangers national security” or induce minors to imitate “unsafe behaviours.” In the context of Chinese regulations, “unsafe behaviour” can also mean supporting Western values like independent journalism or civil society.

So What? Regulating AI-based communication is a top priority for the CAC. While this regulation does help clean up chatbots, it also gives users in China less control over how an AI chatbot speaks to them. That power is now centralized in the hands of platforms.

Share

From Shanghai, With Love

The World AI Conference (WAIC) saw China bolster its ambitions to push its AI products and governance frameworks into the Global South.

In his keynote speech on July 17, Xi Jinping announced that China would launch 5,000 new AI-related training opportunities for developing countries. He also announced China would launch “AI Application Cooperation Centers” in partnership with multilateral intergovernmental organizations, including the African Union, the Arab League and CELAC, whose combined member states cover almost the entirety of the Global South.

China has already established such centers with ASEAN and BRICS. The former is based in Malaysia, a joint operation between a Chinese provincial government and a local company that has government connections and a historical reputation for corruption. The center provides a showcase for Chinese AI products and standards and a springboard for closed-door lobbying by the Chinese government.

The previous day, foreign minister Wang Yi presided over the launch of the World AI Cooperation Organization, floated by premier Li Qiang at last year’s WAIC. According to People’s Daily, the organization proposes to help countries in the Global South develop their AI capacity, to create an open-source ecosystem, and share best practices in AI governance.

The WAIC also included a High-Level Meeting on the Global Governance of AI, attended by foreign dignitaries. The meeting’s official readout recommended a raft of AI policies China adheres to, including preventing AI from being used by “terrorist, extremist and transnational organized crime groups,” and to “promote the adaptation of AI technology to diverse cultural scenarios.”

So What? China already has fingers in several pies on international AI development, but movements at the WAIC signal an expansion in their aims for forging AI in their image in the Global South. While the inaugural members of the WAICO are not the most powerful countries in the world, or with the most well-developed AI systems, focusing on that misses the point. This situation gives China even more power in such an asymmetrically-weighted club, ensuring they get to set the agenda.

Share

This is a placeholder for the “Techy Corner” section hed until I think of something I’m actually happy with, sorry about that folks.

Mind the Gap, It’s Closing

On July 17, the UK’s AI Security Institute (AISI), published a study saying open-weight models now trail frontier closed-source models on offensive cyber capabilities by as little as four months, narrowing the gap of “six to ten months” they estimated last year. The AISI is a body under the British government, internationally-renowned for the quality and impartiality of its AI safety research.

Both of the open-weight models they tested are Chinese. Zhipu’s GLM-5.2, the most cyber-capable open model in the test, matched closed models released four months earlier. DeepSeek V4-Pro matched models five months old — a narrow

[truncated for AI cost control]