Days after the Commerce Department lifted three-week-old export controls on Anthropic's Fable and Mythos AI models over national security concerns, a new report says the company is moving to close loopholes that have allowed Chinese firms to access its most advanced models.
The Financial Times cites people familiar with the matter who say Chinese companies, such as Ant Financial, used overseas subsidiaries, cloud providers, and internal corporate networks to access AI chatbots such as Claude Code.
To note, Alibaba owns about one-third of Ant Financial. Alibaba was recently blacklisted by the US Government over concerns that it was effectively an arm of the Chinese military.
Related:
Ant reportedly provided employees with corporate Claude accounts routed through its Singapore-linked intranet, while ByteDance employees have used VPNs and expense reimbursements for personal Claude subscriptions.
The workarounds do not appear to violate US or Chinese law, but they breach Anthropic's terms of service, which ban Chinese companies and Beijing-controlled foreign entities from using its models.
Anthropic said it prohibits access from unsupported regions, including China, and continuously updates enforcement systems to detect evasion.
FT provided more details on its crackdown:
As part of its efforts to crack down on unauthorised access, Anthropic has targeted "transfer station" services, which relay requests from users in mainland China through Claude accounts registered overseas before returning the responses.
However, larger Chinese AI groups generally avoid transfer stations because the services' operators are widely suspected of storing or reselling prompts. Executives worry rivals could analyse those requests to understand how they are using advanced models to improve their own systems.
What's ironic is that a company behind one of the world's most advanced coding models apparently failed to build the proper guardrails needed to keep Chinese firms from accessing it, even as it works to close those loopholes.