Claude Mythos破解HAWK密碼：AI首次拆穿後量子演算法 | Claude Mythos Cracks HAWK: AI's First PQC Break

7月 29, 2026

By Kit 小克 | AI Tool Observer | 2026-07-29

🇹🇼 Claude Mythos破解HAWK密碼：AI首次拆穿後量子演算法

Claude Mythos Preview破解HAWK後量子密碼演算法的消息，是這週資安圈最熱的話題。Anthropic在7月28日發布研究，證實其尚未公開的Claude Mythos Preview模型，在一個多代理（multi-agent）環境中花費約60小時，獨立找出針對HAWK-256的金鑰還原攻擊，並讓7輪AES-128的破解速度提升200到800倍。這是首次有AI模型在密碼分析（cryptanalysis）領域拿出比人類專家更快的成果，也讓「AI能不能自己做資安研究」這個問題有了具體答案。

Claude Mythos Preview做了什麼？

Claude Mythos Preview是Anthropic內部測試中的模型，尚未對外開放。研究團隊讓它在多代理架構下自主分析HAWK的格（lattice）結構，結果它找到一個先前沒人注意到的數學對稱性，把HAWK-256的預期破解成本從2^64次運算降到2^38次運算——這代表理論上更容易被暴力破解，但仍然是指數級難度，不是多項式時間破解。整個過程只有一位非密碼學專家的人類研究員在旁邊做專案管理，沒有介入實際的數學推導。

HAWK後量子密碼演算法為什麼重要？

HAWK是NIST（美國國家標準與技術研究院）後量子簽章演算法的候選之一，目的是在量子電腦成熟後，取代現行容易被量子演算法破解的RSA、ECC等加密系統。因為攻擊發生在候選階段而非正式部署的系統上，Anthropic也強調這個結果不影響任何生產環境，也不能直接套用到其他格密碼系統。

這對一般開發者有什麼意義？

如果你不是密碼學研究員，這則新聞的重點不是「HAWK被破解了」，而是AI模型分析複雜數學結構的能力正在快速逼近甚至超越人類專家。Anthropic透露整個實驗的API成本大約10萬美元，對一般團隊來說門檻依然不低，但成本會隨模型效率提升持續下降。密碼學社群過去對AI輔助研究抱持懷疑，這次事件很可能加速AI被納入正式的資安審查流程。

好不好用，試了才知道。

🇺🇸 Claude Mythos Cracks HAWK: AI's First PQC Break

Claude Mythos Preview cracking part of the HAWK post-quantum signature scheme is the biggest AI security story this week. Anthropic published research on July 28, 2026 showing its unreleased Claude Mythos Preview model spent roughly 60 hours in a multi-agent setup independently deriving a key-recovery attack against HAWK-256, plus a 200x to 800x speedup on an existing attack against 7-round AES-128. It is the first documented case of an AI model outpacing human cryptographers on a live cryptanalysis problem, and it puts a concrete data point behind the question of whether AI can do original security research.

What did Claude Mythos Preview actually find?

Claude Mythos Preview is an internal, not-yet-released Anthropic model. Working autonomously in a multi-agent environment, it spotted a previously unexploited mathematical symmetry in HAWK lattice structure, dropping the expected cost of breaking HAWK-256 from 2^64 operations to 2^38. A single human researcher, who was not a lattice-cryptography specialist, provided only project-management oversight — the mathematical derivation was the model own work.

Why does the HAWK algorithm matter?

HAWK is one of NIST post-quantum signature candidates, designed to replace RSA and ECC once quantum computers become powerful enough to break them. Because this attack targets a candidate scheme rather than a deployed system, Anthropic was explicit that the result does not affect any production cryptography and does not generalize to other lattice-based schemes.

Does this matter if you are not a cryptographer?

The headline is not "HAWK is broken" — it is still exponentially hard to break, not a polynomial-time break. The real story is that AI is closing the gap on tasks that used to require years of specialized human expertise. Anthropic put the API cost of the whole experiment at around $100,000, still steep for most teams, but that number will keep falling as models get more efficient. Expect this to accelerate how seriously security researchers treat AI-assisted cryptanalysis going forward.

好不好用，試了才知道 — worth testing before you trust the hype.

Sources / 資料來源

常見問題 FAQ

Claude Mythos Preview是什麼？

是Anthropic尚未公開發布的內部測試模型，專門用於高難度推理與研究任務，目前不對外開放使用。

HAWK後量子密碼真的被破解了嗎？

沒有完全破解。這次只是把理論攻擊成本從2^64降到2^38次運算，仍屬指數級難度，不是實際可行的多項式時間破解。

一般人現在使用的加密系統會受影響嗎？

不會。Anthropic明確表示這個結果不影響任何生產環境的加密系統，也不能直接套用到其他格密碼演算法。

這代表AI已經能取代密碼學家了嗎？

還不行。整個過程仍需人類研究員在旁監督專案進度，且結果目前只適用於HAWK和特定AES變種，無法直接通用到其他系統。

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