Claude浮水印評測：應付歐盟新規，一個參數就破解 | Claude AI Watermark Review: One Parameter Bypasses It

11月 02, 2026

By Kit 小克 | AI Tool Observer | 2026-11-03

🇹🇼 Claude浮水印評測：應付歐盟新規，一個參數就破解

Claude浮水印從11月3日開始正式上線：Anthropic為了符合歐盟AI法案（EU AI Act）規定，替四款較舊的Claude模型加上內建浮水印，用來標記內容是AI生成的。但資安研究人員馬上發現，其中兩款模型的浮水印只要調一個取樣參數就能整組失效——合規歸合規，防偽力道卻沒那麼硬。

Claude浮水印是什麼，為什麼現在才補上

這次補上浮水印的四款模型是Claude Mythos 5、Opus 4.7、Opus 4.6、Sonnet 4.6。浮水印的做法不是加看得到的符號，而是在生成文字時微調「選字時的隨機性」，讓輸出帶有統計上可偵測的規律，就算複製貼上到別的地方也會跟著走。Anthropic表示，8月2日之後發布的新模型（像Opus 5.5）出廠就內建浮水印，這次只是把舊模型補齊。

背後的推手是歐盟AI法案第50條：要求AI生成內容必須標記，新系統從2026年8月2日就要做到，已經上線的舊系統則有到12月2日的過渡期。Anthropic今年7月簽了歐盟「AI生成內容透明度行為準則」，等於是提前卡位應付12月大限。

浮水印漏洞：調一個參數就能繞過

問題出在Opus 4.6和Sonnet 4.6這兩款模型上——浮水印只在預設取樣參數下才會啟動。開發者只要透過API調整temperature、top_p、top_k三者中任何一個，浮水印就整個失效，而且不需要什麼特殊技巧，改一個數值就夠了。

對一般用戶和開發者分別意味著什麼

一般聊天用戶：大多數人不會去碰取樣參數，浮水印基本上還在，偵測機制對「沒刻意防範」的內容仍然有效。
透過API開發的人：只要調一個常見參數，就能無痛跳過標記，無論是不是故意，結果都一樣。
想靠浮水印抓AI造假的人：別太樂觀，這從來不是防偽終極方案，頂多抓到不特別防範的那群人。

老實說，Claude浮水印目前做到的是「打勾式合規」——滿足法規字面要求，但防偽強度留了一個明顯後門。如果你是靠浮水印判斷文字是不是AI寫的，這篇文章就是提醒：別把它當萬能解，尤其面對懂技術、想規避的人，一個參數就能讓它形同虛設。

好不好用，試了才知道。

🇺🇸 Claude AI Watermark Review: One Parameter Bypasses It

Claude's AI watermark went live for four older models on November 3, 2026, as Anthropic races to meet the EU AI Act's deadline for marking AI-generated text. The rollout covers Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 — but security researchers immediately found that on two of those four models, the watermark can be switched off by tweaking a single sampling parameter.

What Claude's Watermark Actually Does

Instead of inserting visible symbols, the watermark nudges the randomness behind word selection during generation, leaving a statistically detectable pattern baked into the text itself. Since it lives in the words rather than extra characters, it survives copy-paste into other documents. Anthropic says every model shipped after August 2, 2026 (like Opus 5.5) already launched with the watermark built in — this update just catches up the older lineup.

The driver is Article 50 of the EU AI Act, which requires AI-generated content to be labeled: new systems had to comply from August 2, 2026, while systems already on the market get a transition window until December 2, 2026. Anthropic signed the EU's Code of Practice on Transparency of AI-Generated Content back in July, positioning itself ahead of that December cutoff.

The Loophole: One Parameter Turns It Off

Here is the catch: on Opus 4.6 and Sonnet 4.6, the watermark only activates under default sampling settings. Any developer calling the API can disable it entirely by adjusting temperature, top_p, or top_k — no exotic jailbreak required, just a routine generation parameter most API users already touch.

What This Means in Practice

Casual chat users: most people never touch sampling parameters, so the watermark still applies and detection works as intended for them.
API developers: one routine parameter change — intentional or not — silently removes the mark.
Anyone relying on the watermark to catch AI-written text: temper expectations. It was never meant to stop a motivated actor, only to flag content from people not actively trying to hide it.

Bottom line: Claude's watermark is compliance-by-checkbox — it satisfies the letter of the EU AI Act but leaves an obvious bypass for anyone who knows where to look. If you are counting on it to reliably separate AI text from human text, do not. One parameter tweak is all it takes.

好不好用，試了才知道 / The only way to know if it holds up is to test it yourself.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士