Anthropic has implemented watermarks for Claude

Anthropic has begun implementing text watermarks for four Claude models. The innovation is linked to preparations for the entry into force of the requirements of the European EU AI Act in December.

However, testing revealed a vulnerability in the protection of some early versions. On systems such as Claude Opus 4.6 and Claude Sonnet 4.6, official API users can completely disable watermark generation by simply changing one sampling parameter in the request.

In response to the discovered loophole, the system developers limited their reaction to a formal warning about compliance with regulatory rules. No technical blocks have been introduced at the level of older architectures yet.

The existence of a simple way to bypass the labeling calls into question the compliance of outdated versions with the strict requirements of European legislation for corporate clients.