Anthropic quickly removed a tracker that was secretly monitoring Cloud Code users in China. This issue surfaced after a security researcher exposed hidden spyware-like code called "prompt steganography," describing the action as a serious breach of user trust.
Last week, a web developer named Thereallo encountered code while investigating privacy issues in Cloud Code that covertly sent Chinese users' information, including time zone, proxy, and possible connections to Chinese AI labs. Although the code was not malicious, it sent data to Anthropic without users' knowledge.
Anthropic engineer Tariq Sheppard confirmed on social media platform X that this tracker had been active experimentally since March, aiming to prevent abuse by unauthorized vendors and protect the model from unauthorized "knowledge extraction." Reports indicate that unauthorized vendors sell access to free models cheaper, and professional subscriptions costing up to $100 per month may be offered at much lower prices.