Anthropic removed hidden tracking code from Claude Code in China

In short: Anthropic removed hidden tracking code from its Claude Code tool after a researcher found it was quietly flagging some users in China.

What happened

A security researcher and web developer who goes by “Thereallo” said they found hidden code inside Claude Code, Anthropic’s tool for helping developers write and edit software with AI. The code was not described openly to users, and it was designed to identify signals that a user might be in China.

The researcher said Anthropic used “prompt steganography,” which is a way to hide extra instructions inside the text sent to an AI (like slipping a note into a letter where most people will not notice it). The hidden instructions reportedly caused Claude Code to send Anthropic information that could help flag a user’s time zone, whether they were using a proxy (a service that can hide where you are), and possible connections to Chinese AI labs.

An Anthropic engineer, Thariq Shihipar, confirmed on X that the tracker was added in March as an “experiment.” He said it was meant to reduce account abuse by unauthorized resellers and to protect against “distillation,” which is when one AI system is trained by repeatedly querying another AI, like copying answers from a study guide at massive scale. After the disclosure, Anthropic removed the code.

Why it matters

Developer tools like Claude Code often have deep access to your work, including files and commands on your computer. Even if the tracking was limited, hiding it can weaken trust, especially as AI companies face pressure to stop copying of their models while still respecting user privacy.

Source: Arstechnica