Anthropic 文本水印在多款旧 Claude 模型上可被轻松绕过
Anthropic's text watermark is easy to bypass on several older Claude models
- 1Anthropic 为满足欧盟 AI 法案 12 月期限,分批为旧款 Claude 模型补装文本水印,且客户无法关闭
- 2Claude Opus 4.6 与 Sonnet 4.6 的水印仅在默认采样参数下生效,改动 temperature、top_p、top_k 任一参数即可绕过
- 311 月 16 日起第二轮补装覆盖 Opus 4.6、Opus 4.5、Sonnet 4.5、Haiku 4.5,四款均受默认参数限制
- 4Anthropic 称水印不影响质量速度且不含用户信息;John Gruber 等质疑按水印选词会降低用词精度
文本水印正在变成嵌入模型层且不可关闭的合规默认项,但旧款模型上其有效性取决于采样参数是否保持默认。
- 影响:依赖水印做内容合规的团队需确认采样参数未被调过,否则旧模型上的水印会静默失效;输出可被检测为 AI 生成,禁止 AI 的客户若发现水印可能引发纠纷;官方称定价与接口格式不受影响。
- 参考:可考虑盘点自家调用是否调整过 temperature、top_p、top_k,先算清水印在关键场景是否真正生效;合规上依赖水印的团队值得把参数锁定为默认值,并申请检测 API 私测资格。
- 依据:参数限制与不可关闭等细节出自 Anthropic 致 API 客户的官方邮件,官方口径暂无独立验证;水印无损质量的说法存在 John Gruber 等公开质疑。
Update, November 4, 2026:
Anthropic emailed API customers again about a second round of watermark retrofits. Starting November 16, 2026, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5, and Claude Haiku 4.5 will all get text watermarking. On Amazon Bedrock, the rollout may take a few extra days.
Claude Opus 4.6 was originally part of the first wave and should have gotten the watermark on November 3. Anthropic explicitly confirms the delay to November 16 but gives no reason. The other three first-wave models, Claude Mythos 5, Claude Opus 4.7, and Claude Sonnet 4.6, don't show up in the new email so their rollout likely went ahead on November 3 as planned.
Claude Opus 4.5, Claude Sonnet 4.5, and Claude Haiku 4.5 are three models that weren't part of the previous announcement. All four models in this wave share the same sampling restriction: the watermark only applies when temperature, top_p, and top_k are set to their defaults. In the first wave, that limit only hit two of four models (Opus 4.6 and Sonnet 4.6).
Original article, October 31, 2026:
Anthropic is expanding its text watermark to four more Claude models, but API users can dodge the mark on two of them by tweaking a single sampling parameter.
In an email to API customers, Anthropic said it will watermark Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Newer models like Claude Opus 5.5 already carry the mark, and the rollout on Amazon Bedrock may take a few extra days.
Anthropic calls the watermark an "imperceptible pattern in word choice" that doesn't change meaning, quality, or readability. The company says it adds no extra characters or tokens, doesn't affect pricing or request and response formats, and barely slows output speed. Anthropic also says the watermark "contains no information about the user, their organization, or their conversations with Claude."
Anthropic applies the watermark only at default settings on two older models
On Claude Opus 4.6 and Claude Sonnet 4.6, Anthropic says the watermark kicks in only when developers leave the three sampling parameters at their default values. Those parameters, temperature, top_p, and top_k, control how much randomness the model uses when picking words. A minor tweak to any one of them is enough to dodge the watermark. Anthropic warns in its email that "If you rely on the watermark for your own compliance, keep these settings at their defaults."
The email doesn't mention that caveat for Claude Mythos 5 or Claude Opus 4.7, which suggests the watermark applies to those models regardless of sampling settings. But the loophole on the older models is still notable, since Opus 4.6 remains Anthropic's best writing model, ahead of even Opus 5.5.
Anthropic has a detection API in private preview for verifying the watermark and offers more details on its content labeling information page.
Anthropic retrofits older Claude models to meet the EU AI Act deadline
Anthropic points to the EU AI Act's December deadline as the reason it's watermarking older models now. Every model released after Aug. 2, 2026, shipped with the watermark on day one, while older models are getting it in stages. Since the watermark is embedded at the model level, it works on every platform that hosts these models, including Amazon Web Services, Google Cloud, and Microsoft Foundry. Customers can't turn it off.
Anthropic uses a variant of Google DeepMind's SynthID text method, which tweaks the randomness behind word selection to produce a statistically detectable pattern. OpenAI launched its own watermarking system textGrain in October 2026, but it only activates by default for ChatGPT and Codex users in the EU. API customers worldwide can flip it on or off, and it ships disabled. Google Deepmind open-sourced SynthID for text and built it into the Gemini models.
Whether watermarking truly leaves text quality untouched is disputed. Blogger John Gruber argues that no two synonyms mean exactly the same thing, so picking words by a watermark key instead of semantic fit inevitably degrades precision. The legal trade outlet Artificial Lawyer raises a separate concern for law firms, where clients who've banned AI use could discover the mark, and multiple watermarks layered inside one document could create complications.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
来源:The Decoder:AI News · the-decoder.com