Security researchers uncovered a hidden tracking mechanism within Anthropic's Claude Code, an AI tool designed to assist developers with coding tasks. The code, described as similar to spyware, used a technique called 'prompt steganography' to embed signals within the system. This mechanism was reportedly intended to monitor users in China and gather information such as time zones and proxy usage.

Anthropic engineer Thariq Shihipar acknowledged the tracker was an 'experiment' added in March to prevent unauthorized account abuse and protect against 'distillation attacks', a process where one AI model's output is used to train another. While Anthropic stated the code was meant for deletion and that stronger protections were in place, the stealthy implementation drew significant criticism from privacy advocates and led to Alibaba banning its employees from using Claude Code.

The incident places Anthropic in a difficult position, balancing the need to protect its AI models from imitation against maintaining user trust. The use of such covert methods in a tool that interacts closely with user files and commands raises significant privacy questions, particularly given the sensitive nature of AI coding assistants.