Claude’s China ban has become a market for access—and for users’ data

The Information reports that six tenants in one roughly 30-business Beijing office building sell access to Claude and other US models.

WIRED describes a wider trade in overseas accounts and intermediary services that persists despite Anthropic’s restrictions on access from China. A customer can obtain functioning inference without knowing who controls the accounts or handles the traffic. Working access therefore establishes very little about the trustworthiness of the supplier.

Anthropic’s September threat-intelligence report identifies a second business behind some of these connections. The company says third-party resellers collect user exchanges without consent and sell them to laboratories seeking material for model distillation: training another system on a more capable model’s outputs. It says the collected conversations can also contain corporate information and personal details. These are findings reported by Anthropic, which has commercial and security interests in stopping the trade; they do not establish that every intermediary sells customer data.

The practical implication extends beyond geographic restrictions. A company assessing an inference provider needs to know which organizations can see its prompts, code and responses, how long they retain them and whether reuse is permitted. A low token price omits the potential cost of transferring confidential work to an operator whose incentives and identity are unclear.