Anthropic Removes Hidden Tracking Code in AI Tool Amid Privacy Concerns

Anthropic, an AI company, faced criticism after it was revealed that its programming tool, Claude Code, contained a hidden mechanism designed to detect attempts to bypass the company's security systems. According to reports by c14 and findings shared by developer Therealo, the tool monitored the URL used to access Anthropic's API and, upon detecting an unusual address, collected additional user environment data such as the computer's time zone and hostname. The purpose of this mechanism was to identify users potentially trying to circumvent protections, gain unauthorized commercial access, or train competing models using Claude's outputs.

Developers criticized the lack of clear disclosure to users about this tracking feature, raising concerns about transparency, privacy, and trust in AI development tools. Following the exposure, Anthropic quietly removed the controversial code in an update. Company engineer Tarik Shihipar confirmed on X that the mechanism was an experimental feature initiated in March to combat "model extraction" and misuse of Claude accounts. Model extraction involves external parties feeding large query volumes into commercial models like Claude to train rival models, posing a significant threat to AI companies' intellectual property and technological advantage.

Anthropic stated that more advanced protection mechanisms have since been developed, making the older code obsolete and slated for removal. This incident highlights the growing tension in the AI field between protecting costly models and maintaining transparency with users and developers.