Anthropic's research team has published findings demonstrating that its Claude model was used to develop a full key-recovery attack on the HAWK-256 post-quantum signature scheme and to dramatically accelerate analysis of a reduced-round version of AES-128. The work, detailed in an Anthropic research post, showcases AI models functioning as active drivers of cryptanalytic discovery.
The breakthrough against HAWK-256—a lattice-based signature scheme under consideration in post-quantum standardization—exploits a previously unexamined symmetry in its underlying mathematical structure. Anthropic has released the implementation for this end-to-end attack, which has an expected runtime of about three hours and 42 minutes on a standard 96-core server. This transforms a theoretical vulnerability into a practically demonstrable weakness requiring serious consideration.
Alongside this, the same research effort delivered a significant performance gain for a known attack pathway against seven-round AES-128. The model-assisted method achieved a 200- to 800-fold speedup. It is critical to note that this result applies only to the reduced-round variant; the full ten-round AES-128 algorithm remains secure against this specific approach. Nevertheless, the substantial improvement underscores the potential for AI to uncover new efficiencies in analyzing well-studied cryptographic primitives.
The dual outcomes—one impacting a next-generation post-quantum proposal, the other enhancing an analysis of an established block cipher—illustrate the broad scope of AI-assisted cryptanalysis. Traditionally, finding such non-obvious mathematical relationships required intense, specialized human effort. The new capability appears to lie in the model's proficiency at generating complete, functional attack chains.
The release of practical attack code aligns with Anthropic's stated approach to transparency in responsible disclosure. By providing working implementations, the company enables independent verification and further scrutiny from the security community, which is vital for integrating these advanced methods into rigorous, ongoing algorithm evaluation.
For organizations and standards bodies navigating the transition to quantum-resistant cryptography, these findings add a crucial data point. They suggest that candidate evaluation must be a continuous process, subject to evolving analytical techniques. Similarly, the AES result reinforces that even long-deployed standards benefit from periodic reassessment with modern tools.
Important questions remain for further study, such as what structural changes would neutralize the exploited symmetry in HAWK-256 without compromising its efficiency or security claims. The scalability of these AI methodologies to full-round versions of major ciphers is another open area. The research signals a move toward establishing frameworks for systematic, ethical AI-driven security auditing.
The core takeaway is clear: large language models have progressed to a point where they can generate original, actionable discoveries in highly specialized domains. As these tools mature, their responsible integration into cryptographic review processes will be a defining challenge for the security community, aimed at strengthening the foundations of digital trust.
Anthropic研究團隊發表了研究成果，證明其Claude模型被用於針對HAWK-256後量子簽名方案開發完整的金鑰還原攻擊，並大幅加速了對簡化輪次AES-128的分析。這項工作在Anthropic的研究文章中詳細闡述，展示了人工智能模型如何作為密碼分析發現的主動驅動力。
針對HAWK-256（一種正在後量子標準化進程中考量的格基簽名方案）的突破，利用了其底層數學結構中一個先前未經檢驗的對稱性。Anthropic已釋出此端對端攻擊的實作方案，該攻擊在標準96核伺服器上的預計運行時間約為3小時42分鐘。這將一個理論漏洞轉化為實際可示範的弱點，需要嚴肅考量。
與此同時，同一研究項目在針對七輪AES-128的已知攻擊路徑上取得了顯著的性能提升。模型輔助方法實現了200到800倍的速度提升。必須注意的是，此結果僅適用於簡化輪次變體；完整的十輪AES-128演算法針對此特定方法仍然安全。儘管如此，巨大的改進突顯了人工智能在分析已被充分研究的密碼學原語時，發掘新效率的潛力。
這兩項成果——一項影響下一代後量子提案，另一項增強了對已確立區塊密碼的分析——展示了人工智能輔助密碼分析的廣泛範疇。傳統上，發現此類非顯見的數學關係需要高度專業化的人類密集努力。新的能力似乎在於模型生成完整、功能性攻擊鏈的熟練程度。
釋出實用攻擊代碼與Anthropic在負責任揭露中所聲明的透明度方針一致。透過提供可用的實作方案，公司使安全社群得以進行獨立驗證和進一步審查，這對於將這些先進方法整合到嚴格、持續的演算法評估中至關重要。
對於正在過渡到抗量子密碼學的組織和標準制定機構而言，這些發現增加了一個關鍵數據點。它們表明候選方案的評估必須是一個持續的過程，需適應不斷演進的分析技術。同樣地，AES的結果強化了即使是長期部署的標準，也能受益於使用現代工具進行定期重估。
仍有重要的問題有待進一步研究，例如進行何種結構性改變可以在不影響HAWK-256效率或安全聲明的前提下，中和被利用的對稱性。這些人工智能方法學在主要密碼的完整輪次版本上的可擴展性，是另一個開放領域。這項研究標誌著朝向建立系統化、合乎道德的人工智能驅動安全審計框架的發展。
核心要點很明確：大型語言模型已發展到一個程度，能夠在高度專業的領域產生原創、可行的發現。隨著這些工具的成熟，它們被負責任地整合到密碼學審查流程中，將成為安全社群面臨的決定性挑戰，旨在鞏固數位信任的基礎。