Claude AI Just Cracked a Post-Quantum Test Scheme and Found a Faster 7-Round AES Attack Claude AI 刚刚破解了后量子测试方案，并发现了更快的7轮AES攻击

TL;DR

• Claude Mythos Preview successfully derived an end-to-end key-recovery attack against HAWK-256, a post-quantum signature scheme candidate in NIST's standardization process.
• The attack exploits a previously unused symmetry in the lattice structure of HAWK, reducing the security level from 2^64 to 2^38 for HAWK-256 and improving gate-count estimates for larger parameters.
• For AES-128, Mythos enabled a 200- to 800-fold speedup on a seven-round variant by removing a 256-way guessing step via a novel invariant fingerprint called the Möbius Bridge.
• Neither result affects production systems: HAWK-256 is a challenge parameter only, and the AES attack requires impractical numbers of chosen plaintexts.
• Anthropic released full technical papers, reproducibility artifacts, and code, with Mythos conducting most research autonomously under human oversight.

Why It Matters

This breakthrough demonstrates that advanced AI agents like Claude Mythos can independently discover non-trivial cryptographic vulnerabilities in standardized schemes, potentially accelerating cryptanalysis but also raising concerns about AI-assisted attacks on future standards. For practitioners, it underscores the need for rigorous, automated verification of cryptographic designs and highlights how AI may reshape both offensive and defensive capabilities in cybersecurity. Researchers should consider these findings when evaluating the long-term resilience of post-quantum candidates and symmetric ciphers under evolving analytical tools.

Technical Details

• HAWK-256 Attack: Exploits a hidden automorphism (symmetry) in the lattice underlying the HAWK signature scheme. Constructs a τ-cocycle lattice from the public key, applies lattice reduction and sieving to recover short vectors, then reconstructs a functionally equivalent signing basis (592-byte decoded key). Runtime estimated at ~3h 42m on a 96-core server; reduces work factor from 2^64 to 2^38 for HAWK-256. Gate counts improved for HAWK-512 (2^150 → 2^108) and HAWK-1024 (2^288 → 2^182), though these remain infeasible.
• Seven-Round AES-128 Attack: Modifies meet-in-the-middle attacks by introducing the "Möbius Bridge"—an invariant fingerprint unaffected by a 256-way enumeration step previously required. This eliminates redundant computations, yielding 200–800× speedup depending on metric. Requires ~2^105 chosen plaintexts, rendering it impractical for real-world use. Code includes C/Python/Rust components for validation but does not execute full end-to-end recovery on actual AES-128.
• AI Role: Mythos Preview operated autonomously over ~60 hours in a multi-agent setup, with humans providing only project direction, compute resources, and verification. API cost was ~$100,000. All results are reproducible via published artifacts.
• Scope Limitations: HAWK attack targets only HAWK-256 (challenge parameter); AES attack applies to reduced-round variants and remains computationally prohibitive at scale. No impact on full-round AES or other NIST signature candidates.

Industry Insight

Cryptographic standardization bodies like NIST must now incorporate AI-driven cryptanalysis into their evaluation frameworks, as autonomous agents can uncover subtle structural weaknesses invisible to traditional methods. Organizations adopting post-quantum algorithms should prioritize schemes with conservative security margins and undergo independent red-teaming using advanced AI tools before deployment. Meanwhile, developers of symmetric ciphers should design resistance against invariant-based optimizations early in the specification phase to preempt similar speedups in reduced-round analyses.

TL;DR

• Claude AI 的 Mythos Preview 成功破解了后量子密码 HAWK-256 方案，将密钥恢复的工作因子从 $2^{64}$ 降低至 $2^{38}$。
• 该攻击利用格密码中此前未被利用的对称性（automorphism），结合格规约和筛选技术实现端到端密钥恢复。
• 在 AES-128 七轮变体上，通过引入“莫比乌斯桥”不变指纹，移除了原有中间相遇攻击中的 256 步猜测步骤，加速 200–800 倍。
• 两项成果均未影响生产系统：HAWK-256 仅为 NIST 挑战参数，AES 攻击需极大量选择明文且未覆盖完整 10 轮。
• Anthropic 强调人类提供方向与验证，AI 自主完成核心推导，API 成本约 10 万美元，耗时约 60 小时。

为什么值得看

本文展示了大模型在密码分析领域的实质性突破能力，尤其是对复杂数学结构（如格对称性、代数不变量）的自动发现与利用，标志着 AI 正从辅助工具转向主动科研参与者。对于安全从业者而言，这既是警示——传统密码设计可能低估 AI 对隐式模式的挖掘潜力；也是机遇——可借助类似机制加速漏洞审计与新方案评估。

技术解析

• HAWK-256 攻击路径：基于 van Gent & Pulles 提出的非平凡自同构简化 smLIP 问题，Mythos 进一步发现额外对称性，构建 τ-cocycle 格，经 lattice reduction 和 sieving 找回短向量，最终重构功能等价签名基（592 字节），而非原始密钥种子（96 字节）。实际运行于 96 核服务器预计耗时 3h42m。
• AES-128 七轮优化：针对 meet-in-the-middle 攻击中需枚举 256 个中间值的瓶颈，Mythos 设计出“Möbius Bridge”——一种在该变换下保持不变的指纹特征，从而跳过穷搜环节。虽未执行完整 10 轮恢复，但组件测量外推显示整体效率提升显著。
• 实验复现性：Anthropic 发布两篇论文及代码 artifact，包含 C/Python/Rust 多语言实现模块，并对小型 AES-like 24bit 密钥完成黑盒全恢复，确保结果可验证。
• 安全性边界澄清：明确区分“挑战参数”与“标准参数”，HAWK-512/HOWK-1024 的 gate-count 仅小幅下降（如 HAWK-512 从 $2^{150} \to 2^{108}$），仍属不可行范围；AES 攻击所需 $2^{105}$ 选明文远超现实场景。
• 人机协作模式：研究者非密码学专家，仅提供项目管理与算力支持，AI 独立承担公式推导、算法构造与逻辑闭环，体现当前 AGI 雏形下的科研范式迁移。

行业启示

• 后量子密码标准化进程需纳入 AI 风险评估：NIST 应建立针对 AI 驱动 cryptanalysis 的动态审查机制，尤其在第三轮候选方案中增加对潜在对称性或代数结构的自动化压力测试。
• 加密工程实践应预留“AI 安全余量”：在设计新协议时，除考虑经典计算复杂度外，还需预估模型是否能识别隐藏不变量或低维投影，建议采用更高冗余度或随机化扰动策略。
• AI 赋能的安全审计将成为常态：企业可部署类似 Mythos 的内部代理系统定期扫描自身加密实现，提前暴露弱配置；同时警惕第三方依赖库是否被用于生成对抗样本或侧信道线索。