Anthropic's watermark rollout leaves a loophole on two older Claude models
Reported by The Decoder1 source ↓
Anthropic is adding text watermarks to four Claude models ahead of the EU AI Act's December deadline. On Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by changing one sampling parameter.
Key points
- The rollout covers Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6
- On Opus 4.6 and Sonnet 4.6, changing temperature, top_p or top_k can bypass the mark
- Anthropic cites the EU AI Act's December deadline for adding watermarks to older models
- Amazon Bedrock's rollout may take a few extra days
Details
Anthropic told API customers it will add watermarks to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. The Decoder reports that the watermark is an imperceptible word-choice pattern and does not add characters or tokens. Anthropic says it does not affect pricing or request and response formats.
For Opus 4.6 and Sonnet 4.6, the watermark applies only when temperature, top_p and top_k remain at their default settings; changing any one can bypass it. Anthropic advised customers who rely on the watermark for compliance to keep those settings at their defaults. The company's email did not describe the same caveat for Mythos 5 or Opus 4.7.
Anthropic says the EU AI Act's December deadline prompted the rollout to older models; models released after Aug. 2, 2026, have had watermarks from launch. The watermark is embedded at the model level and works across hosting platforms, including Amazon Web Services, Google Cloud and Microsoft Foundry. Anthropic's detection API is in private preview, and the company says customers cannot turn the watermark off.
Written from coverage as of
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh