Anthropic has taken steps to close avenues through which some Chinese companies accessed its Claude language models and coding tools in breach of the company’s restrictions on China, the Financial Times reported.
What methods were used?
Sources say several Chinese firms, including Ant Financial, relied on a range of workarounds. These included using cloud providers and overseas subsidiaries: companies created enterprise Claude accounts that employees could reach over the firm’s internal networks or via a Singapore subsidiary. ByteDance — the parent company of TikTok — does not appear to provide direct Claude access, but this year introduced a reimbursement scheme under which engineers can reclaim the cost of personal subscriptions and frequently use the tools via virtual private networks (VPNs).
Why are these workarounds an issue?
Although such arrangements may not contravene US or Chinese law, they violate Anthropic’s terms of service, which explicitly prohibit use of the models by Chinese companies and their overseas subsidiaries. Anthropic has enforced stricter controls: it requires firm user identification and blocks transactions originating from Chinese banks.
How does Anthropic’s approach compare with others?
The report notes that Chinese users can in practice gain easier VPN-based access to OpenAI tools because OpenAI does not apply equally stringent checks. Anthropic’s coding tools such as Claude Code are particularly popular among Chinese software developers and AI startups because generated outputs are often reused for model distillation — training smaller models to imitate the behavior of more advanced systems.
What concrete steps has Anthropic taken?
Anthropic has moved to curb routes that operated via Microsoft Azure and overseas subsidiaries. Access provided through the Microsoft API to entities registered in Singapore enabled engineers on the Chinese mainland to reach Claude over internal networks. Microsoft said Anthropic, with Microsoft’s support, closely monitors service usage and enforces the terms of use.
Anthropic has also acted against intermediary proxy servers that forward requests from the Chinese mainland through Claude accounts registered abroad. Larger Chinese AI firms tend to avoid such proxy services because they fear operators might store or sell incoming data.
Monitoring signals and other checks
Earlier, Anthropic used Claude Code to search for telltale signals — for example, a computer’s time zone — that could indicate a user is connecting from China. Such techniques are part of the company’s broader effort to ensure compliance with its rules and terms.
Why this matters
The measures underscore tensions that arise where US AI exporters’ restrictions intersect with global user demand and technical means to bypass controls. Anthropic’s strict stance and technical enforcement illustrate how US companies are attempting to uphold their policies in environments where cloud services, VPNs and overseas affiliates can be used to circumvent limits.
Tags: China, Microsoft, Anthropic, artificial intelligence, cloud services, Azure, OpenAI, VPN, ByteDance, software development, generative AI
