Claude破解NIST曲線評測:25年密碼學懸案30分鐘解開 | Claude NIST Curve Review: 25-Year Crypto Mystery Solved in 30 Min
By Kit 小克 | AI Tool Observer | 2026-10-17
🇹🇼 Claude破解NIST曲線評測:25年密碼學懸案30分鐘解開
密碼學界流傳了25年的都市傳說終於有解:1990年代NSA交給NIST的NIST橢圓曲線(P-192到P-521,也就是今天TLS、SSH、各種數位簽章底層常見的曲線)裡,那組神秘的「種子」數字到底怎麼來的?NSA當年沒給任何解釋,只丟出一串16進位數字,讓後來的研究者懷疑裡面是不是藏了後門。這個懸案在2026年10月被Claude在不到30分鐘內解開了。
Claude做了什麼
根據Anthropic官方研究部落格,一位內部員工丟給Claude一個問題:你能不能自己想辦法查出這些種子是怎麼生成的?沒有給任何提示。Claude猜測這些種子可能來自1990年代C語言標準庫常見的lrand48偽隨機數產生器,再用時間戳記當初始值去反推——這其實是一個標準的格基(lattice)問題。結果Claude不只猜對了演算法,還把P-256曲線的種子精確還原回1997年8月的一個時間點,其他幾條NIST橢圓曲線(P-192、P-224、P-384、P-521)的種子來源也全部對上。
為什麼這件事重要(但沒有你想的那麼嚴重)
這不是一次新的密碼學攻擊,也沒有發現NIST橢圓曲線有漏洞。恰恰相反——Anthropic指出,種子的隨機性空間只有大約30位元,範圍太小,不足以讓人刻意挑選出一條「弱」曲線藏後門。換句話說,Claude的發現反而是替NIST橢圓曲線「洗白」:25年的陰謀論懷疑,終於有了清楚、可驗證的技術解釋。
有人獨立驗證過,不是AI自己說了算
這點才是整件事最值得學的地方:密碼學研究者Filippo Valsorda(Let's Encrypt背後ISRG的成員)用Go語言重新實作了一遍Claude的推論過程,結果完全吻合,還把原本的懸賞獎金加碼到36,864美元,全數捐給ISRG。延伸閱讀可以看Dev.to的技術拆解文章。AI給出的結論不管聽起來多厲害,都要有獨立人工驗證才能當真,這次剛好是一個教科書級的範例。
老實說,這對你有什麼用
- 如果你不是密碼學研究者,NIST橢圓曲線目前不用換、不用緊急修補,這則新聞不會改變你明天的工作。
- 它實際示範了一種用法:把Claude當「研究助理」去啃一個長年沒人解開的技術謎團,確實可行,而且快到不合理(30分鐘)。
- 前提永遠是找得到獨立驗證的方法,而不是看到AI給答案就直接信。
同一批研究裡,CoinDesk的報導也提到Claude被拿去重新設計量子電路,某些版本的閘數減少到原本的1/17。邏輯是一致的:AI加速探索,人類負責把關驗證。
好不好用,試了才知道。
🇺🇸 Claude NIST Curve Review: 25-Year Crypto Mystery Solved in 30 Min
Cryptographers have argued for 25 years about where the seed values behind the NIST elliptic curves (P-192 through P-521 — the curves underpinning TLS, SSH, and most ECDSA signatures today) actually came from. The NSA handed NIST a block of hex digits in the 1990s with zero explanation, fueling decades of backdoor speculation. In October 2026, Claude cracked it in under 30 minutes.
What Claude Actually Found
According to Anthropic's research blog, an employee simply asked an internal Claude model to figure out how the seeds were generated — no hints given. Claude guessed the seeds came from lrand48, a pseudorandom number generator shipped with 1990s C standard libraries, seeded with a Unix timestamp, then recovered the generator's internal state by treating it as a textbook lattice problem. It nailed every curve: P-256's seed decodes to a specific moment in August 1997, and P-192, P-224, P-384, and P-521 all matched the same pattern.
Why It Matters (and Why It's Less Dramatic Than It Sounds)
This is not a new attack and it doesn't break the NIST elliptic curves. If anything, it's good news: the seed's randomness space is only around 30 bits — far too narrow for anyone to have cherry-picked a deliberately weak curve. Claude's discovery replaces 25 years of backdoor paranoia with an actual, checkable technical answer.
Independently Verified — Not Just AI Saying So
The part worth remembering: cryptographer Filippo Valsorda, a member of ISRG (the nonprofit behind Let's Encrypt), independently reimplemented Claude's reasoning in Go and confirmed the exact same match, then tripled the original bounty to $36,864 and donated it to ISRG. See the write-up on Dev.to for the technical reimplementation details. That's the real lesson: an AI's conclusion, however impressive, only counts once a human re-derives it independently — and this time it did.
The Honest Takeaway
- If you're not a cryptographer, nothing changes for you tomorrow — the NIST elliptic curves don't need patching or replacing.
- This is a real, practical use case: pointing an LLM at a decades-old unsolved technical puzzle and letting it grind through the math in half an hour.
- The catch is always the same — you need a human expert to independently verify the answer, not just trust the AI's output.
The same research cycle also covered Claude redesigning quantum circuits, per CoinDesk's report, with one variant using 17x fewer gates. Same pattern: AI accelerates the search, humans still do the verification.
好不好用,試了才知道。
Sources / 資料來源
- Anthropic Research: Claude reveals ECDSA curve origins
- Dev.to: Claude Traced the NIST Curve Seeds to a 1997 Clock and lrand48
- CoinDesk: Claude cracks a 25-year NSA curve mystery
延伸閱讀 / Related Articles
- Anthropic AI代理評測:工時超員工6倍,初階職缺剩9% | Anthropic AI Agent Review: 6x Staff Hours, Junior Hiring Crash
- Apple HomeView評測:299美元智慧顯示器登場,歐盟缺席 | Apple HomeView Review: $299 Smart Display, No EU Launch
- OpenAI Dots評測:GPT-6 Astra常駐代理,CAPTCHA卡關 | OpenAI Dots Review: Always-On Agent Stumbles on CAPTCHAs
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
留言
張貼留言