China moves to control AI model exports, a free agent jail, and Taiwan's Texas chip parks

China's Ministry of Commerce and Ministry of Science and Technology published a draft on Friday, October 16, that would add 40 technologies to the country's catalogue of technologies prohibited or restricted from export. More than half of them, 23, are in artificial intelligence. The rest are five biomedicine technologies and 12 in photovoltaics. According to Crypto Briefing's reading of the draft, 14 of the 40 are marked prohibited and 26 restricted. The public can comment until November 15, after which Beijing will publish a final version.

The AI list reportedly covers large AI models, AI agents, AI chip design, computing clusters, voice recognition, high-performance translation, deepfake detection and hacking-related technologies. The most striking change is that model weights and source code come into scope for the first time, but only for products that exceed certain performance thresholds. Those thresholds have not been published in the summaries available so far. At the same time, the ministries introduced an interim registration requirement covering all the listed technologies, which one analyst quoted by The National said applies immediately, including to exports already under way.

This did not come out of nowhere. In July, the Financial Times and Reuters reported that the commerce ministry had been consulting Alibaba, ByteDance, Zhipu and other firms about limiting overseas access to China's most advanced models. Chinese labs dominate the open-weight leaderboards, and developers worldwide, including many in Africa, build on models such as Qwen, DeepSeek and GLM. The timing is also political: the draft lands weeks after a US-China trade truce struck in September, and Wendy Cutler of the Asia Society Policy Institute told The National that the 'fragility of our trade truce cannot be overstated.'

There are big open questions. Reporting by TokenPost and others says access through APIs, websites, apps and software-as-a-service platforms would be exempt, as would technology already public through open-source releases or academic papers. If that holds, today's downloadable Chinese models are unaffected; the question is what happens to future releases above the thresholds. Leonhard Xu of the consultancy Sinolytics stressed that 'export control does not equal an export ban in most cases' and that the announcement is so broad it will take time to assess. This is a draft, and the final text can change.

The practical effect for builders is a new kind of supply risk. A team that self-hosts a Chinese open-weight model is fine today, but cannot assume the next, stronger version will be published openly. Companies doing cross-border licensing, acquisitions or IP transfers with Chinese partners may now need to register those deals. And the move strengthens the case for model-agnostic software: products that can switch between Chinese, US and European models without a rewrite are less exposed when any government changes the rules.

Why it mattersChinese open-weight models power a large share of low-cost AI products worldwide. If future frontier releases are controlled, builders who rely on downloading and self-hosting them face a new supply risk.

Perplexity published the code for SPACE on October 15. It is the sandbox platform that has run every session of the company's Computer agent since July, and Perplexity says it handles millions of sandbox creations a day. The code is on GitHub under the permissive Apache 2.0 licence, which means anyone can use it, change it and build commercial products on it. Three cloud providers are already on board: Akamai has launched it on its cloud platform, Crusoe is adding it to its managed AI services, and CoreWeave has published setup guidance for its Kubernetes service.

The design is simple to explain. Each agent gets its own small virtual machine, a Firecracker microVM with its own operating system, so code the agent runs is walled off from the host computer. A control plane creates, pauses, resumes and deletes these machines. SPACE treats every sandbox as untrusted: network access and credentials are handled outside it, so an agent can reach approved services without ever holding the keys. Snapshots and checkpoints let a session be restored or branched cheaply. It is written in Rust, and a compatibility wrapper lets some code written for the commercial sandbox service E2B run on it.

Perplexity's security claims are its own. In its blog post the company says its evaluation found no escape from the virtual machine to the host in 108 attempts; Semafor and SiliconANGLE report that nine frontier models made 118 attempts to break out and none fully escaped. Some testers did find ways around network restrictions, which Perplexity says it fixed before follow-up tests. According to SiliconANGLE, the same tests broke several proprietary sandboxes. Perplexity also says SPACE created sandboxes three to five times faster than its previous provider on the same production traffic.

The context is a run of agent-containment failures. Semafor notes that AI labs have reported agents escaping their constraints and reaching systems such as Hugging Face and Australian government websites, and the FTC is already investigating AI agent incidents. Nvidia released a similar open-source sandbox, OpenShell, last month. These results have not been independently audited, the two reported attempt counts differ, and the repository only lists partial E2B compatibility. Running SPACE yourself also means operating Kubernetes and virtual machines, which is real infrastructure work.

The bigger shift is that strong agent isolation is becoming free. Until now, small teams building agents that run code or browse the web either paid a sandbox vendor or ran agents in loosely locked-down containers. A battle-tested, permissively licensed alternative lowers that cost and raises the baseline that clients will expect. It also creates work: someone has to deploy it, configure network rules and credentials, test it and keep it updated, and most small businesses deploying agents will not do that themselves.

Why it mattersSafe isolation is the main thing standing between a useful code-running agent and a security incident. A free, production-tested option makes it realistic for small teams to do this properly.

Taiwan plans to build a series of industrial parks in the United States to help its smaller AI hardware suppliers set up near their biggest customers, Economy Minister Kung Ming-hsin told Bloomberg in an interview. The first would be in Texas. Bloomberg reported the plan this week, and it was picked up widely on October 15 and 16. Kung estimated each complex would cost $2 billion to $4 billion, with the first costing up to $4 billion. The aim is to bring the parts makers that feed TSMC and server assemblers such as Foxconn, and through them Nvidia, onto US soil.

The money would not come mainly from Taiwan's state budget. The Taiwan Electrical and Electronic Manufacturers' Association, an industry body, is partnering with the government and is expected to raise the funds. The government's role is support: loan guarantees and help navigating local and federal US regulations, and, according to 24/7 Wall St., help with visas and land access. The model is clustering. Instead of each small supplier taking on the cost and risk of a US factory alone, they would share a site, services and paperwork.

There is already a template in Texas. Wistron opened a $700 million plant in the state that assembles Nvidia AI systems, and has said a second facility is in the works. TSMC has pledged $265 billion for US expansion, mostly in Arizona, and its management said on an October 15 call that Texas is one option for more capacity. The plan also serves politics: Taiwan faces pressure to meet investment commitments under its trade deal with Washington, and it wants to reduce supply-chain exposure to tensions with China.

The caveats are substantial. Nvidia, TSMC and Foxconn have made no commitment to the parks, according to reporting, and no participating suppliers, schedule or customer contracts have been named. TSMC said it has not announced any investment in Texas. One report cited people familiar with the matter saying land had been secured, but the ministry did not confirm that, and the economy ministry said it would share details 'when the timing is right'. Building factories in the US is also slower and more expensive than in Taiwan, which is exactly why small suppliers have held back.

If the parks go ahead, they would move a slice of the world's AI hardware supply chain, the cables, cooling parts, power supplies, boards and metal work around every AI server, closer to US data-centre demand. That means years of construction, hiring, logistics, compliance and services work around Texas. For people outside the US, the more realistic openings are remote: research, translation, documentation, procurement support and digital services for the hundreds of mid-sized Taiwanese firms that will have to make this move.

Why it mattersAI server production depends on hundreds of small suppliers, not just chipmakers. Moving them as a cluster would make US AI hardware supply faster and less exposed to geopolitical risk, if it happens.