AI Watermark Remover Hits 11k Stars in 5 Days Defeating Anthropic's Claude Provenance

An MIT licensed toolkit designed to strip provenance markers from AI generated content saw a significant increase in adoption on GitHub following the release of Anthropic's new text marking system. The project reached approximately 11,000 stars within five days, targeting watermarks from Claude, SynthID-Text, and OpenAI across files such as PNG, PDF, and DOCX. It specifically identifies and removes three levels of provenance, including invisible Unicode characters, statistical token-sampling patterns, and C2PA and EXIF metadata.

The tool arrived shortly after Anthropic detailed the implementation of Claude's global watermarking to comply with the EU AI Act. The software is built as an agent skill that operates a Python service over HTTP to perform deterministic scripts on Unicode and model guided rewrites for statistical marks. Claude currently refuses to install the watermark removal plugin.