Anthropic watermark rollout leaves loophole on two older Claude models
Original title: Anthropic's watermark rollout leaves a loophole on two older Claude models
API users can bypass watermark on Claude Opus 4.6 and Sonnet 4.6 by changing sampling parameters.
Anthropic expands text watermark to four Claude models but API users dodge it on two older ones via tweaks to temperature, top_p or top_k. The company retrofits older models to meet the EU AI Act deadline. Watermark uses a Google DeepMind SynthID variant and applies at model level across hosting platforms.
Unconfirmed: Whether watermarking truly leaves text quality untouched is disputed.
- Watermark applies to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6
- Loophole exists on Opus 4.6 and Sonnet 4.6 at default sampling settings
- Watermark adds no extra tokens or affects pricing or speed
- Detection API available in private preview
- Vendors
- Anthropic
From the original article
2 min read at the-decoder.com
Anthropic is expanding its text watermark to four more Claude models to meet the EU AI Act's December deadline. On the older models Claude Opus 4. 6 and Sonnet 4.
By Matthias Bastian · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.