Anthropic's Claude Text Watermark Easily Removed via API Parameter
Anthropic is rolling out text watermarking on four more Claude models, but older models such as Claude Opus 4.6 and Sonnet 4.6 can be bypassed via a single sampling parameter.
Anthropic is bringing text watermarking to four more Claude models. At the same time, it has emerged that the marking can be circumvented with little effort on several older models. Affected are, among others, Claude Opus 4.6 and Claude Sonnet 4.6. API users can remove the watermark via a single sampling parameter.
The use of the watermarks is connected to the requirements of the EU AI Act, whose provisions take effect in December. The obligation to label AI-generated content is the main driver of this step. The report leaves open whether the bypass option meets compliance requirements.
Anthropic has responded to the finding with a note on compliance with the rules, but no fix for the affected models has been announced. It remains unclear whether the company will subsequently secure the older models. It is also unclear whether versions beyond the two named models are affected.
The actual extent of misuse of the bypass has not yet been documented. For providers of generative AI, the question is becoming more important of how reliable provenance markings remain technically when regulation mandates them.
Source: the-decoder.com
4
