Fable 5.1密碼評測：370年謎題破解惹爭議 | Fable 5.1 Review: Cipher Solved, Scholars Push Back

9月 15, 2026

By Kit 小克 | AI Tool Observer | 2026-09-15

🇹🇼 Fable 5.1密碼評測：370年謎題破解惹爭議

Fable 5.1（Claude Fable 5.1）在 2026 年 9 月宣稱解開了一個懸而未決 370 年的密碼——蘇格蘭作家 Thomas Urquhart 在 1653 年著作中留下的「Cyphral Distich」。這是本週 Hacker News 熱度最高的 AI 話題（單篇破千點讚），但隔天就有學者提出質疑，是一則很適合拿來檢視「AI 研究成果該怎麼驗證」的案例。

Fable 5.1 怎麼破解這個密碼的？

這個密碼是一串 64 個數字，1899 年就被列在《Notes and Queries》期刊上公開徵解，後來也被密碼學者 Klaus Schmeh 列進「50 大未解密碼」名單，代表不是隨便什麼工具都能碰的老問題。根據 Vals AI 公開的測試紀錄，Fable 5.1 在完全沒有人工介入的情況下，花了 44 分鐘、消耗約 17.6 萬 tokens 找出解法。

關鍵不是暴力破解，是找對「金鑰」

Fable 5.1 沒有用傳統的頻率分析或替換式解密
它推論出「金鑰就是這本書本身」——用密碼裡的數字去對應 Urquhart 書中自己標示的 32 個段落（Proquiritations）
取每個對應段落的第一個字母，串起來就是一句祝福查理二世復位的押韻對句
模型同時嘗試了更長、285 個數字的變體版本，但因為原始文本抄寫不一致，留下 9 個位置無法確定

為什麼學者隔天就潑冷水？

資安評論人 Bruce Schneier 在部落格追蹤此事時指出關鍵爭議：有研究者翻查大英圖書館館藏的 1653 年版本後發現，那個版本裡根本沒有這串密碼，只有裝飾花紋和結語。密碼其實出現在其他版本（1652 年格拉斯哥版、1834 年 Maitland Club 版），但不是每個版本都有，實體比對還在進行中，尚未逐字元完成獨立驗證。

對一般使用者的實際意義

Fable 5.1 展現的是長時間、無人監督的搜索與驗證能力：面對結構化的可能性空間，它能自己嘗試假設、驗證、排除，這對寫程式除錯、資料分析這類任務很有參考價值。但這起事件也提醒我們，AI 給出的「解答」聽起來多有把握，都需要跟人類研究一樣接受同儕檢驗——尤其是牽涉歷史文獻版本這種需要實體查證的領域，模型再厲害也無法取代原始文獻比對。

好不好用，試了才知道。

🇺🇸 Fable 5.1 Review: Cipher Solved, Scholars Push Back

Fable 5.1, Anthropic's Claude Fable 5.1, made headlines this month for reportedly cracking a 370-year-old cipher — the Cyphral Distich, buried at the back of Scottish writer Thomas Urquhart's 1653 book. It topped Hacker News with over a thousand points, but the celebration didn't last: researchers pushed back within a day, making this a good case study in how to actually verify an AI research claim instead of just applauding it.

How Did Fable 5.1 Solve the Cipher?

The cipher is a 64-number cryptogram that's been a known open problem since it was posed in Notes and Queries in 1899, and it later made cryptologist Klaus Schmeh's list of the top 50 unsolved historical ciphers — not a toy puzzle. According to a test published by Vals AI, Fable 5.1 solved it in 44 minutes using about 176,000 tokens, with no human intervention during the search.

The Trick Wasn't Brute Force — It Was Finding the Right Key

Fable 5.1 skipped standard frequency analysis or substitution-cipher approaches entirely
It reasoned that the key is the book itself, using each cipher number to index into one of the 32 sections (Proquiritations) Urquhart explicitly labeled in his own text
Taking the first letter of each indexed word produced a 32-letter rhyming couplet wishing King Charles II a successful reign
The model also attempted a longer, 285-number variant, but left 9 positions unresolved due to transcription inconsistencies across surviving copies

Why Did Researchers Push Back the Next Day?

Security researcher Bruce Schneier flagged the core dispute on his blog: when someone checked the British Library's actual 1653 copy, the cipher wasn't there at all — just decorative flourishes and closing text. The cipher does appear in other editions (an 1652 Glasgow copy, an 1834 Maitland Club edition), but not consistently across all surviving copies, and full character-by-character independent verification against physical copies was still in progress.

What This Actually Means for Everyday Use

What Fable 5.1 demonstrated is genuinely useful: sustained, unsupervised search-and-verify over a structured possibility space, the same skill that matters for debugging code or digging through messy datasets. But the episode is also a reminder that a confident-sounding AI answer deserves the same scrutiny as a human research claim — especially in a field like historical textual criticism, where no model can substitute for checking the physical source material.

好不好用，試了才知道。

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