Anthropic’s watermark loophole lets users bypass marks on two older Claude models
Anthropic is rolling out text watermarks to four Claude models but admits users can bypass them on two older versions by adjusting a single sampling parameter. In an email to API customers, the company announced watermarks for Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. However, the watermark on Opus 4.6 and Sonnet 4.6 only activates at default settings for…
Key points
- Anthropic’s watermark bypasses on Claude Opus 4.6 and Sonnet 4.6 by tweaking sampling parameters
- EU AI Act deadline drives retrofitting of older models, but newer ones had watermarks from launch
- Critics argue watermarks may degrade text quality or complicate legal compliance
Anthropic cites the EU AI Act’s December deadline as the reason for retrofitting older models. The watermark uses a variant of Google DeepMind’s SynthID method, embedding an imperceptible pattern in word choice without altering meaning, speed, or cost. Critics, including blogger John Gruber and Artificial Lawyer, question whether watermarks degrade text quality or create legal complications, such as revealing banned AI use in client documents.
Model page: Claude Opus 5.5 →
The story so far
3 episodes →- Anthropic’s watermark loophole lets users bypass marks on two older Claude modelsthis story
Anthropic's watermark rollout leaves a loophole on two older Claude models
The Decoder · 31 October 2026
Loading the full article…
This text was published by The Decoder and written by Matthias Bastian. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Policy & Regulation
All →- OpenAI, Google, Microsoft, and Anthropic will honor a shared robots.txt signal for AI training preferences · 9 src
- Albanese faces backlash from Grayndler artists over AI scraping proposal · 1 src
- Anthropic’s Claude blocks game cloning and reverse-engineering, users report · 2 src
- OpenAI to train models on Daybreak Blue data from Nov 29 · 1 src
- DOJ busts $200M student loan AI fraud ring in Operation Ghostbusters · 1 src
Comments
via GitHub Discussions