Anthropic的Claude Code被发现在系统提示里藏了检测中国代理的暗码，虽然已经回滚，但做法不太透明。看看他们怎么解释的。
研究人员发现Claude Code在系统提示中通过不可见Unicode字符编码用户时区和代理信息，用于检测中国代理和Asia/Shanghai、Asia/Urumqi时区。Anthropic员工证实这是2026年3月的实验，并已在被发现前开始回滚。争议焦点在于方法而非检测意图——使用混淆代码而非透明遥测渠道。该实验旨在识别转售和蒸馏行为，但回滚已在今天发布中合并并推送。
News has been going around for a few days about Cl…
News has been going around for a few days about Claude Code allegedly hiding China proxy detection in system prompts. Researchers found code that encoded user timezone and proxy info into invisible Unicode characters.
This was a March 2026 experiment that Anthropic staff say was already being rolled back before anyone discovered it. The code checked for Chinese proxies and Asia/Shanghai or Asia/Urumqi timezones, then modified date formatting and swapped invisible apostrophes to signal detection.
The issue isn't that they tried to detect resellers and distillation attempts. Other AI labs do telemetry too. The problem was the method - encoding signals in prompt content using obfuscated code rather than documented telemetry channels users could inspect or opt out of.
Anthropic staff confirmed they'd built stronger mitigations since March and were planning to remove this anyway. The rollback was merged and shipped in today's release. https://t.co/ALputSjqmg