Anthropic's watermark rollout leaves a loophole on two older Claude models

What happened

Anthropic is expanding text watermarking to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6, but changing temperature, top_p, or top_k bypasses it on the latter two models. Its customer email does not mention that caveat for Mythos 5 or Opus 4.7. Anthropic cites the EU AI Act's December deadline for the rollout and says the watermark adds no characters or tokens and contains no user or conversation information.

Read next

• More changes in today’s briefingConfigure the ChatGPT desktop app with Amazon Bedrock in AWS GovCloud (US)

The story so far

Anthropic told API customers it will expand text watermarking to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6, but developers can bypass it on the latter two by changing sampling settings. For Claude Opus 4.6 and Claude Sonnet 4.6, Anthropic says watermarking applies only when temperature, top_p, and top_k all remain at their default values; even a minor change to any one disables it. The customer email does not state the same caveat for Claude Mythos 5 or Claude Opus 4.7. Anthropic cites the EU AI Act’s December deadline as the reason for the rollout. The company says the watermark adds no characters or tokens and contains no user or conversation information.