AI Pulse · Under the Radar

Anthropic announced a watermarking scheme for Claude last week. A developer named mikiane shipped a working remover on GitHub the next day. The repo, claude-watermark-cleaner, already has 98 stars and is picking up roughly 72 new ones every 24 hours. It runs one or two rewrite passes on Claude output until about 70% of the token sequences are disrupted, enough to break the watermark. The developer framed it explicitly as proof the system isn’t robust.

A second tool from Jesper Nissen extends the same approach to Gemini and OpenAI watermarks. Announcement to working break took under ten days.

What this means for you

If you’re using AI output in your business, watermarks were supposed to give you a way to prove provenance or flag synthetic content. That window just closed faster than anyone expected. The technical details matter less than the timeline. A motivated person with decent coding skills can now strip these markers in minutes, and the tools are public.

This doesn’t mean watermarking is pointless, but it does mean you can’t treat it as a reliable gate. If you’re building compliance workflows that depend on detecting AI content, or if you’re worried about employees passing off AI work as their own, watermarks won’t solve it. You need process, not just technical flags. That’s the kind of thing we build into an AI command centre, tracking what gets used, by whom, and how it flows through your operation, not just whether a token pattern survived a rewrite.

The broader point: adversarial tools move fast. If a control mechanism can be bypassed in a day, plan around the bypass, not the mechanism. Watermarks might still have a role in low-stakes contexts, but don’t bet your risk framework on them.