Skip to main content
LLMgram · AI News · 2026-10-31

Anthropic extends Claude watermarks for the EU AI Act but Opus 4.6 and Sonnet 4.6 stay bypassable via sampling

Anthropic extends Claude watermarks for the EU AI Act but Opus 4.6 and Sonnet 4.6 stay bypassable via sampling

Anthropic is expanding its text watermark to four more Claude models ahead of the EU AI Act's December deadline, according to The Decoder. The move aims to label generated text for compliance, but the report says Claude Opus 4.6 and Sonnet 4.6 retain a notable gap: developers can bypass the watermark by changing one sampling parameter. Anthropic has addressed that loophole with a compliance warning rather than preventing the tweak, The Decoder reports. For builders and operators, the detail matters because watermarking is often treated as provenance evidence even though caller-controlled generation settings can weaken it. The available excerpts do not specify which parameter enables the bypass, whether newly watermarked models share the flaw, or if Anthropic will patch Opus and Sonnet before regulators expect robust marking.

Sources

Open in AI Signal →

Anthropic extends Claude watermarks for the EU AI Act but Opus 4.6 and Sonnet 4.6 stay bypassable via sampling

Anthropic extends Claude watermarks for the EU AI Act but Opus 4.6 and Sonnet 4.6 stay bypassable via sampling

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 4.6, developers can bypass the watermark by tweaking a single sampling parameter.

Key takeaway

A watermark that yields to one sampling-parameter change is compliance theater, not cryptographic provenance for generated text.

What happened

The Decoder reports that Anthropic is expanding its text watermark to four additional Claude models to meet the EU AI Act's December deadline, extending marking requirements across more of the Claude lineup.

On the older Claude Opus 4.6 and Sonnet 4.6 models, The Decoder says developers can bypass the watermark by tweaking a single sampling parameter, and Anthropic has responded with only a compliance warning rather than blocking that workaround.

Evidence

  • Anthropic is expanding its text watermark to four more Claude models for the EU AI Act December deadline.

    The Decoder · attributed

    Anthropic is expanding its text watermark to four more Claude models to meet the EU AI Act's December deadline.

  • Opus 4.6 and Sonnet 4.6 allow a single-parameter sampling bypass of the watermark.

    The Decoder · attributed

    On the older models Claude Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by tweaking a single sampling parameter.

  • Anthropic's response to the bypass is characterized as a compliance warning only.

    The Decoder · attributed

    Anthropic addresses the loophole with just a compliance warning.

Why it matters

Regulators and enterprise buyers should calibrate EU AI Act watermark expectations against models where API callers can still strip marks trivially.

Limits and uncertainties

The packet excerpts do not name the exact sampling parameter, the technical watermark design, or whether the bypass affects the four newly covered models.

Practical implications

Teams using Opus 4.6 or Sonnet 4.6 for compliance-sensitive outputs should not rely on Anthropic watermarks alone without independent provenance controls.

What to watch

Whether Anthropic closes the Opus 4.6 and Sonnet 4.6 sampling bypass or patches those models before the EU AI Act December deadline.

Sources

LLMgram editorial selection and synthesis · @llmgram. LLMgram is not the original publisher of this information.
Continue on LLMgram: Open in AI Signal →
Original reporting: Anthropic's watermark rollout leaves a loophole on two older Claude models