Anthropic Hid Secret Tracking of Claude Code Users for Three Months

A developer known as Thereallo found a hidden mechanism in the Claude Code binaries that had been quietly profiling users connecting from Chinese time zones for three months, unnoticed. The discovery triggered a wave of criticism against Anthropic, a company that has spent years building its image as an opponent of excessive surveillance of AI users.

The technique Anthropic used resembles classic steganography. Instead of openly collecting telemetry data, the system subtly altered the text of responses generated by the model depending on detected geographic or network signals, in a way that was practically invisible to the average user but could be detected and decoded by the company itself.

How the Hidden Code Worked

The mechanism changed the date format in the system prompt from a hyphenated form to a slash-based form for users connecting from time zones associated with China. At the same time, it swapped the regular apostrophe in the sentence "Today's date is" for a visually indistinguishable character with a different Unicode code point. These microscopic differences in the text acted as a hidden marker, letting Anthropic distinguish and track specific categories of users without disclosing this in its terms of service or changelog.

The discovery would not have been possible without the persistence of a single developer, who noticed unusual inconsistencies in the formatting of Claude Code's responses and began analyzing the tool's binaries. The story quickly spread through the developer community and then reached wider tech media.

The code was an experiment added in March, and we had been planning to remove it for some time - Thariq Shihipar, Anthropic engineer

Anthropic's Explanation and Industry Reaction

According to Shihipar, the mechanism was meant to protect the company from abuse by unauthorized resellers and from distillation, the practice of training competing models on Claude's outputs. That explanation did not satisfy critics, who point out that regardless of intent, hidden and undisclosed profiling of users based on location contradicts Anthropic's stated mission as a company that puts safety and privacy above growth at any cost.

The situation is made more sensitive by the fact that Anthropic had previously refused US government agencies access to Claude for mass surveillance of American citizens, citing privacy principles. Critics point out that the company applied different standards to its own undisclosed analytics mechanisms than to third-party requests.

What It Means for Coding Tool Users

The case damages trust in coding agents at a moment when Claude Code and similar tools are becoming a daily part of developers' work around the world, including in Poland. Hidden analytics mechanisms in tools that have access to code repositories and company data raise questions about what other undisclosed features might be running in popular AI assistants without the knowledge of the teams deploying them in their workflows.

For companies using Claude Code in Poland and Europe, the case means, in practice, a need to review AI tool vendors' privacy policies more carefully and to monitor changes in client-side code, especially where regulations such as GDPR require transparent disclosure of data processing. Anthropic announced it would remove the mechanism but did not disclose the full list of data it collected or the exact period it had been active beyond the three-month window mentioned.

Sources: Decrypt (decrypt.co), Futurism (futurism.com), Slashdot (yro.slashdot.org)