Anthropic Cracks Down on Chinese Workarounds to Claude

Detection Methods: How Anthropic Identifies Unauthorized Access

Anthropic uses a layered technical enforcement system:

• Geolocation blocking and IP proxy detection: Claude directly blocks access from China and requires overseas credit cards for payment ^I.
• Live biometric identity verification: As of April 2026, Anthropic runs real-time passport-plus-selfie checks on new accounts for certain Claude features ^FI.
• Behavioral monitoring: Anthropic spokesperson Michael Aciman says the company uses "a range of measures to identify and block unauthorized access, including tracking proxy usage patterns and other technical indicators" ^I.
• Large-scale account pattern detection: The company actively monitors for clusters of fraudulent accounts (e.g., 24,000+ fake accounts linked to DeepSeek, Moonshot, and MiniMax), which it identifies through usage pattern analysis ^TM.

Chinese Workarounds That Anthropic Has Targeted

Ant Group

The Financial Times reported in July 2026 that Ant Group (an affiliate of Alibaba) has been accessing Anthropic's tools such as Claude Code through unauthorized workarounds — including providing employees with corporate Claude accounts linked to its Singapore-based entity ^FMI.

ByteDance

ByteDance implemented an internal reimbursement scheme that allowed engineers to use personal accounts to access Claude via VPNs, then get reimbursed by the company, bypassing corporate-level restrictions ^FI.

'Transfer Stations' (中转站)

A thriving gray market has emerged in China where operators run "transfer stations" — intermediary services that buy Claude API tokens or Max accounts from outside China and redistribute access domestically ^SLM. Typical scheme: buy one Max 20x account for $200, use Claude Code OAuth tokens to expose it as "unlimited API," then sell access to 10–20 clients at $30–50/month, generating $300–1,000 per account ^I. These stations are marketed openly on GitHub, Taobao, and Telegram, accept RMB payments, and resell access at roughly 10% of the official price ^SLM.

Model Substitution and Prompt Harvesting

Some Chinese proxy services engage in "model substitution," silently routing queries to cheaper or older Claude models while charging for premium ones, and harvesting prompts as a secondary revenue stream ^I.

Escalating Tensions: Accusations Against Chinese AI Labs

February 2026 — DeepSeek, Moonshot AI, and MiniMax

On February 23, 2026, Anthropic publicly accused three Chinese AI labs — DeepSeek, Moonshot AI, and MiniMax — of conducting large-scale "distillation" attacks ^TMD. The companies allegedly created approximately 24,000 fraudulent accounts, generated over 16 million exchanges with Claude, used "hydra cluster architectures" to disguise the activity, and extracted model reasoning and tool-use capabilities to train their own competing AI models ^MD. The Guardian described this as "large-scale intellectual property theft" ^D. OpenAI had made similar allegations against Chinese firms the previous month ^D.

June 2026 — Alibaba

On June 24, 2026, Anthropic accused Alibaba Group of waging a "large-scale effort to illicitly access Claude using thousands of fraudulent accounts" ^F. Alibaba subsequently announced that starting July 10, 2026, it would prohibit employees from using Anthropic's Claude Code in the office environment ^Y. Alibaba also alleged that Anthropic had "embedded a backdoor into Claude Code" — though this claim is unsubstantiated and comes via a third-party report ^Y.

The Ongoing Enforcement Push (July 2026)

As of July 3, 2026, Anthropic is moving to plug the remaining loopholes that allow unauthorized Chinese access, according to the Financial Times ^FM. The company is tightening KYC checks, enhancing proxy detection, and likely revoking accounts tied to known "transfer station" operations. The crackdown follows months of escalating cat-and-mouse dynamics: each time Anthropic adds a new barrier — geoblocking, then credit card requirements, then biometric verification — Chinese gray-market operators find a new workaround ^ISL.