Unicode espionage: Why does the AI ​​assistant Claude Code secretly monitor programmers?

A loud scandal is shaking the global IT community. Console AI assistant Claude Code, which has full access to developers' terminals and local files, has been caught secretly collecting data. Researchers discovered that the utility uses sophisticated steganography inside prompts, secretly transmitting user information to Anthropic servers. While some accuse AI creators of creating spy backdoors, others see it as the heavy hand of US export controls.

The Claude Code tool was positioned as a secure corporate assistant capable of working inside closed circuits, writes xrust. However, an independent audit of version 2.1.91 showed that if you run the utility through a third-party proxy server, it activates a hidden total scan mechanism.

How Unicode masking works

The AI assistant locally checks the user's system settings. If a Chinese time zone is detected (Asia/Shanghai or Asia/Urumqi), or the proxy address matches the list of servers of Chinese IT giants (Baidu, Alibaba, ByteDance), the program begins to secretly encode this metadata.

Instead of sending telemetry openly, Claude Code modifies the standard service line of the system prompt: Today's date is… («Today's date…»).

• Character substitution: The usual apostrophe in the word Today's is imperceptibly replaced by visually identical Unicode doubles (for example, ’, ʼ or ʹ). The combination of these hidden characters serves as a marker for Anthropic's servers.
• Games with dates: For users from the “black list” the date is formatted using a slash (2026/06/30), and for others — with a hyphen.

The language model and the developer himself see plain text, but the Anthropic server side instantly reads these “watermarks”.

Split in the community: Espionage or protection?

The Hacker News and Reddit forums are literally ablaze with controversy. The developers are furious: the utility has Shell access to the system, and hidden modification of prompts undermines any trust. Today AI calculates how to bypass sanctions, and tomorrow it will start scanning private code for “political correctness” or leaking trade secrets.

On the other hand, Anthropic is under severe pressure from US regulators. The company is trying to protect its advanced models from illegal export and “distillation” (when competitors train their neural networks on the responses of someone else’s AI).

The introduction of Unicode steganography is associated with Anthropic’s attempt to prove to the White House its trustworthiness in matters of technology export control, even at the cost of total surveillance of users.

Xrust Unicode espionage: Why does the AI ​​assistant Claude Code secretly monitor programmers?