Anthropic's Watermark Rollout Leaves a Loophole on Two Older Claude Models

dawn.fi

Factuality

Ownership

Artificial intelligence giant Anthropic has informed its customers about a change to the way its Claude language models work that will take effect next week. According to the company, a digital watermark will be added to all Claude models it offers on November 3rd. The watermark will allow text to be identified as being generated using artificial intelligence. Anthropic introduced the watermark embedded in the text back in August for its latest …

DiarioBitcoin

Factuality

Ownership

Anthropic expands its watermark system to four Claude models, but in Opus 4.6 and Sonnet 4.6 a small change in sampling parameters can avoid the signal. The difference matters for developers who rely on this function in their compliance controls.

the-decoder.de

Factuality

Ownership

Anthropic extends its text watermark to four more Claude models, driven by the December deadline of the EU-AI Act. However, with the older models Claude Opus 4.6 and Sonnet 4.6, a small change in the sampling parameters is enough to bypass the marking – a loophole that Anthropic only comments on with a compliance note. The article Anthropic API users can handle watermarks in older Claude models with a trick first appeared on The Decoder.