Summary of Key Points
Recently, Claude, an AI tool developed by Anthropic, has extensively banned Chinese users, employing a highly secretive detection mechanism. It not only circumvents VPNs but also identifies Chinese users by reading the local time zone and using intermediary domain names on their computers. Additionally, it utilizes steganography to make invisible modifications to system prompts, covertly transmitting user identification to servers. This opaque behavior has sparked outrage among developers both domestically and internationally, completely eroding users' trust in Anthropic.
I. The Ban Process: From Email Tracing to Code “Espionage”
In the past couple of days, almost all Chinese Claude users have been banned, and one of the author’s two Max subscription accounts was also terminated. Even more troubling is that Anthropics ban emails contained hidden address tracking features. However, the real revelation came when a Reddit expert reverse-engineered the Claude code and discovered that it functions like spyware, silently collecting local information and transmitting signals indicating whether a user is Chinese to the server.
II. The Logic Behind Bypassing VPNs
Traditional regional blocking relies on IP addresses, but Claude uses two methods to bypass this:
1. Local Time Zone as an Identifier: While you can change your IP address using a VPN, it’s unlikely to change your computer’s time zone (you wouldn’t set it to the US just to use Claude). Claude directly reads your computer’s time zone settings, and if it’s set to Asia/Shanghai or Asia/Urumqi, the user is identified as Chinese.
2. **Intermediary Domain Name “Blocklist”: Chinese users cannot use the official API and must use intermediary servers to send requests to Anthropic. The Claude code contains a list of 147 encrypted domain names (including those belonging to major companies like Meituan, ByteDance, and Baidu). If a user is detected using one of these domains, they are identified as Chinese. This list is encrypted using XOR with the key 91, making it impossible to reverse-engineer.
III. The Cleverest Trick: Invisible “Steganographic Markings”
Once a Chinese user is identified, how does Claude transmit this information? It uses steganography, hiding the identification within seemingly normal system prompts. For example, in a normal prompt that says “Today’s date is 2026-06-30,” Claude makes two modifications that are invisible to the naked eye:
- Unicode Replacement for Single Quotes: The regular single quote (“Today’s”) is replaced with one of three identical Unicode characters.
- Date Separator Change: If the user’s time zone is detected as Chinese, the hyphen (-) in the date is changed to a slash (/), resulting in “2026/06/30.”
These modifications are detectable by machines but completely invisible to humans. Upon receiving the request, the server can determine the user’s nationality and proceed with the ban.
IV. The Crux of the Trust Crisis: High Permissions and Lack of Transparency
Why is this such a problem for developers?
1. Excessive Permissions in Claude’s Code: Claude has access to read your computer files, execute Shell commands, and modify configurations, essentially granting it “superuser” privileges on your device. Secretly adding identification markers is like someone using your keys to leave secret signs on the door that only accomplices can recognize.
2. Violation of Industry Transparency Standards: Regulations like GDPR require clear communication about data collection, and even Anthropics’ own security documentation emphasizes transparency. However, by using steganography and encrypted blocklists, it deceives users completely.
3. Spreading Trust Crises: If this behavior is targeting China today, it could target other countries tomorrow. Developers are unsure of whether there are more hidden detection methods, and no one wants to rely on such an unreliable tool. Trust takes years to build but can be lost in an instant.
V. The Impact of the Incident: Developers’ Resignation and Anthropic’s Reputational Damage
Reddit posts about this incident have received over 1 million views, and developers worldwide are furious. Many are switching to other AI tools, such as GPT-4 or domestic alternatives like DouBao. Anthropics has severely damaged its reputation by banning Chinese users in such a secretive manner. When it launches new tools, users will likely wonder if similar tactics will be used again.
In short, Anthropic resorted to extremely underhanded methods to block Chinese users, only to sink deeper into a trust crisis.
Conclusion: This incident is not just about banning Chinese users; it raises the fundamental question of whether AI companies respect user privacy and transparency. Users may accept being excluded from certain services, but they cannot tolerate being secretly monitored and labeled—this is a clear violation of their basic trust.