Anthropic Ships Watermarks. A Developer Cracks Them in Four Hours. Compliance Is Not a Technical Problem.
Anthropic announced it would embed invisible, machine-readable watermarks into all Claude-generated content globally to comply with new EU regulations. Developer Guillaume Meyer published an override within four hours of the announcement. The EU rules require watermarks in all new AI models released from August and integration into existing models by December.
This illustrates what security researchers call the asymmetry of offense and defense. A watermarking system that took months to design was circumvented in the time it takes to watch two movies. The underlying lesson is that any security measure baked into consumer-accessible software will be reverse engineered faster than it can be updated. Provenance is a policy problem dressed up as a technical one.
Anthropic announced the watermarking initiative to comply with EU regulation. Developer Guillaume Meyer published a working override within four hours of the announcement, demonstrating the fragility of the approach.
- Open Claude or any AI chatbot and generate a paragraph of text.
- Copy that text into a plain text editor and make three small edits, changing a word, removing a comma, and swapping a sentence order.
- Ask yourself whether any invisible watermark in that text would survive your edits. If you are honest, the answer is no. You have just demonstrated, in under five minutes, why this entire regulatory approach is structurally fragile.