Anthropic tightens China access to Claude after workarounds

Anthropic is tightening access to Claude for users in China after repeated workarounds, the Financial Times reported. The move raises the commercial stakes around a US policy fight over who can use the company’s most advanced systems. It also turns a familiar product-security problem into something more expensive: proof that frontier AI groups can enforce the restrictions they announce.

Blocking a country in a product policy is relatively simple. Keeping that block intact after launch is not.

The work sits across export-control policy, compliance engineering and market reach. Anthropic now has to account for proxies, reseller accounts, remote identity checks and users willing to trade verified access in chat groups. Washington’s pressure is showing up in the mechanics of chatbot distribution, not just in the language of official restrictions.

WIRED reported that users in China had been sharing ways around Anthropic’s geolocation and identity barriers. Semafor reported this week that the company had rolled back China tracking code. The two reports described an enforcement system still being adjusted in public, with workarounds that could be copied, refined and sold.

WIRED said Anthropic introduced identity verification for some Claude users in April. Researcher Hieu Minh Ngo told WIRED that Chinese chat groups were openly discussing how to bypass know-your-customer checks and buy verified access to the service.

“They are talking about how to bypass KYC, where to buy the Claude KYC so they can use it.”

Hieu Minh Ngo, quoted by WIRED

That forum activity points to distribution channels, not casual misuse by a handful of users. Once verified access can be bought and resold, a regional block becomes less like a switch and more like an enforcement program.

Reuters reported on June 13 that Anthropic disabled access to its most advanced AI models after a US order limiting foreign access. The episode showed how quickly Washington can redraw the commercial map for frontier models. It also made gaps between a company’s published restrictions and real-world usage harder to treat as routine trust-and-safety problems.

Reuters separately reported on June 24 that Anthropic said Alibaba had illicitly extracted Claude capabilities through 25,000 fraudulent accounts that generated 28.8 million exchanges between April 22 and June 5. That allegation moved the debate beyond retail workarounds. The figures described access abuse large enough to matter for model security, product economics and the credibility of Anthropic’s controls.

Semafor’s report on the rollback of China tracking code pointed to the trade-off Anthropic is trying to manage. Loose controls risk leakage. Heavy controls can catch legitimate overseas users, researchers and enterprise customers in the same net. For a company seeking broad commercial adoption while staying inside US security expectations, the lane is narrow.

Why the clampdown matters

The China issue is no longer only about whether Claude is officially available in a restricted market. Anthropic has to show that access limits can survive launch, payment channels, usage monitoring and account resale. Regulators, cloud partners and paying customers all want proof that restrictions attached to advanced systems can hold outside a press release.

Export-control pressure is moving deeper into the software layer. Chips and formal access rules were the first fight. The next one looks more like continuous compliance: who is verified, how accounts are watched, how fast loopholes are closed and how much friction a provider can impose before it slows growth. Anthropic’s latest clampdown suggests compliance engineering is becoming part of the product itself.

That lesson reaches beyond Anthropic. As US authorities draw a tighter perimeter around advanced AI access, model providers will be judged on their ability to stop restricted users from circling back through secondary channels. Claude’s China loopholes are also a stress test for the business model behind global AI distribution.