Claude破解NIST曲線評測:25年懸案,30分鐘解開 | Claude ECDSA Review: 25-Year NIST Mystery Solved in 30 Min
By Kit 小克 | AI Tool Observer | 2026-10-20
🇹🇼 Claude破解NIST曲線評測:25年懸案,30分鐘解開
Claude(Anthropic的AI模型)這週登上各大科技媒體頭條,原因不是又刷新了什麼跑分,而是它在30分鐘內解開了困擾密碼學界25年的NIST曲線種子之謎。這個謎題跟Bitcoin、HTTPS、SSH背後都在用的ECDSA加密標準直接相關,連密碼學研究者Filippo Valsorda開出的12,288美元懸賞都因此找到了主人。
事情是怎麼發生的
1990年代NSA公布了P-192、P-224、P-256、P-384、P-521這幾條NIST橢圓曲線的「隨機」種子數字,卻從未說明這些數字從何而來。這留下一個長年的安全疑慮:萬一NSA故意挑了有弱點的種子,等於在全世界的加密系統裡埋了後門?Google、OpenSSL、Bitcoin的簽章演算法ECDSA全都建立在這些曲線上,謎底揭曉前,誰都無法百分之百排除風險。
Claude做了什麼
Anthropic員工出於好奇,直接讓內部研究模型去查這些種子的來源。Claude掃過34種1990年代常見的隨機數產生器組合,鎖定C語言老函式庫裡的lrand48,再用LLL格基簡化演算法(lattice reduction)反推出產生器的內部狀態,整個過程花不到30分鐘。結果發現:P-256的種子是1997年8月16日一次lrand48執行中第1,842個輸出,其餘幾條曲線的種子時間都落在1997年5月到1998年4月之間——說穿了就是拿當時的系統時間戳記當種子,沒有任何神秘操作。
對加密安全意味著什麼
- 這不是破解了ECDSA本身,曲線安全性完全沒變
- 長年的「NSA蓄意留後門」疑慮大幅降低,因為種子生成方式單純到任何人都能重現
- B-163的種子來源,以及為什麼候選值876被跳過,目前依然沒有答案
老實說
這件事真正值得關注的不是密碼學本身解謎,而是AI研究代理處理這類橫跨歷史文獻、統計特徵比對、數學反推的複合任務時展現出的效率——人類密碼學家花了25年沒解開,Claude半小時找到答案。但也別神化,這題目本質上是「系統化排列組合+已知演算法」,不是什麼創造性突破,換成夠有耐心的研究團隊一樣做得到,只是沒人願意花那個時間。對一般用戶來說,這跟你的HTTPS連線安不安全沒有直接關係,但確實是AI輔助研究正在加速解決「堆積已久但沒人想碰」技術債難題的一個真實案例。
🇺🇸 Claude ECDSA Review: 25-Year NIST Mystery Solved in 30 Min
Claude, Anthropic'''s AI model, made headlines this week not for a new benchmark score, but for cracking a 25-year-old cryptography mystery in under 30 minutes: the origin of the NIST curve seeds behind ECDSA. The puzzle sits directly under Bitcoin, HTTPS, and SSH, and it mattered enough that cryptographer Filippo Valsorda'''s $12,288 bounty finally has a winner.
What actually happened
In the late 1990s, the NSA published "random" seed values for the P-192, P-224, P-256, P-384, and P-521 NIST elliptic curves without ever explaining where the numbers came from. That silence fueled two decades of speculation: what if the NSA had cherry-picked weak seeds to plant a backdoor in curves used by Google, OpenSSL, and Bitcoin'''s own ECDSA signatures? Without an answer, nobody could fully rule it out.
How Claude solved it
An Anthropic employee got curious and simply asked an internal research model to figure out where the seeds came from. Claude scanned 34 combinations of random-number generators common in 1990s software, zeroed in on the old C library function lrand48, then used LLL lattice reduction to recover the generator'''s internal state — all in under 30 minutes. The answer: the P-256 seed was the 1,842nd output from an lrand48 run started on August 16, 1997, and the other curves''' seeds trace to Unix timestamps between May 1997 and April 1998. In short, someone just seeded a PRNG with the system clock.
What it means for crypto security
- This doesn'''t break ECDSA — the curves''' security math is unchanged
- The long-standing NSA-backdoor theory loses most of its footing, since the seed generation turns out to be simple and fully reproducible
- Some loose ends remain: B-163'''s seed is still unexplained, and nobody knows why candidate #876 for P-256 was skipped
The honest take
The interesting part isn'''t the crypto trivia — it'''s what this says about AI agents grinding through messy, multi-step research: historical pattern-matching, statistical fingerprinting, and lattice math, stitched together end to end. Human cryptographers had 25 years and didn'''t crack it; Claude needed half an hour. But don'''t overhype it either — this was systematic brute-force search over known algorithms, not a creative leap. Any sufficiently patient research team could have done the same; nobody just bothered. It won'''t change whether your HTTPS connection is safe today, but it'''s a real example of AI research tools chewing through old, neglected technical debt that humans kept putting off.
Sources / 資料來源
- Anthropic 官方研究說明:Claude reveals ECDSA curve origins
- Hacker News 討論串:Claude reveals ECDSA curve origins
- Cyber Security News 報導:Claude AI Solves 25-Year NSA Cryptography Mystery
延伸閱讀 / Related Articles
- Timestral評測:Mistral免GPU時間序列預測模型 | Timestral Review: Mistral's Open CPU Forecasting Model
- Muse Spark Open評測:Meta開源816B模型,美系最強開權重 | Muse Spark Open Review: Meta's 816B Open-Weight Model
- Microsoft AI代理評測:內部威脅論,60款工具現形 | Microsoft AI Agent Review: Insider Threat, 60 Tools Found
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