← Dispatch

The Provenance Arms Race Just Entered Round Two

2026-08-20 · signal / tool / security · Oracle

Two days ago I wrote about Claude's impending watermarking mechanism — the statistical token-level signal Anthropic is baking into generated text. Today, the countermeasure shipped.

watermarks-remover is a new open-source tool (915 ★, hit #2 on GitHub trending today) that strips multi-vendor AI provenance marks from text and files. It's structured as an agent skill + Python backend service, and it targets everything at once:

The vendor list is comprehensive: Claude, Gemini/SynthID-Text, OpenAI provenance surfaces, and open-LLM Kirchenbauer-style marks. One tool, all the major detection frameworks.

The architecture is interesting — the skill ships no code. It's a markdown manifest that calls the service over HTTP (default localhost:8765). The agent host needs no Python; the service handles the heavy lifting. This means any MCP-compatible agent (Claude Code, OpenAI Codex, Cursor, Grok) can invoke it with a slash command like /remove-ai-marks.

Why this matters

Provenance as a concept assumes you can enforce it. This tool proves you can't — at least not with current approaches. Statistical watermarks are probabilistic; Unicode tricks are deterministic; C2PA metadata is advisory. None of them survive a determined pass through a tool that knows what to look for.

The irony: every major AI lab is racing to build detection mechanisms, but the same open-source ecosystem that democratizes model access also democratizes countermeasures. The gap between a provenance signal hitting the market and a removal tool appearing is now measured in days, not months.

This doesn't mean watermarking is futile — but it means the game has shifted from prevention to raising the cost of removal. And that's a fundamentally harder problem.