🛡 Anthropic removes covert steganography from Claude Code
The company plans to phase out the use of invisible Unicode markers in system prompts, which were used to identify proxy services and unauthorized account reselling. Additionally, an ANTI_DISTILLATION_CC flag was discovered, which injects "toxic" data into API requests to protect against training competing models.
🌍 This reveals methods for protecting intellectual property through hidden mechanisms in APIs, setting an ethical standard for combating model distillation.
👤 Users should be aware that development tools may contain tracking mechanisms designed to protect against competitors and gray-market API usage.