Anthropic利用其Claude Mythos模型发布了两项密码分析研究成果。[1]其中一项是针对后量子签名方案HAWK的新型密钥恢复攻击，另一项则是对简化版AES的改进攻击。[1]
新的HAWK攻击将该方案的安全性比特数约减半，理论上可通过加倍密钥长度来修复。[1]该攻击代码已能在数小时内运行于削弱版挑战实例，并已生成可验证的实际代码。[1]相比之下，AES攻击需要2^89次密码操作和2^105次选定明文加密，与2013年的先前工作相比仅为常数倍改进。[1]
这些研究成果的关键价值在于展示了AI在综合应用现有密码学工具方面的能力。[1]Claude Mythos能够理解现有密码分析结果、将其综合并扩展，而非发明全新数学原理。[1]其中HAWK攻击因可能影响标准化进程而更具意义。[1]
验证已成为新的瓶颈。虽然HAWK攻击可直接验证，但AES等改进攻击的有效性检验需要形式化证明或专家评审。[1]
Anthropic has released two cryptanalysis research breakthroughs produced by its Claude Mythos model: a novel key recovery attack on the post-quantum signature scheme HAWK, and an improved attack against simplified AES [1]. The HAWK attack reduces the security bits by approximately half, a weakness that could theoretically be remedied by doubling the key length [1]. The attack code runs within hours on weakened challenge instances and has generated verifiable working code [1].
The second contribution involves improvements to attacks on simplified AES, requiring 2^89 cryptographic operations and 2^105 chosen-plaintext encryptions, representing only a constant-factor improvement over prior work from 2013 [1]. According to analysis of these results, the significance lies in AI's capacity to understand existing cryptanalytic findings, synthesize them, and extend them rather than invent novel mathematics [1]. The HAWK attack is considered more consequential, potentially affecting standardization processes, while the AES attack constitutes incremental progress [1].
A new bottleneck has emerged in validating these cryptanalytic advances [1]. While the HAWK attack permits direct verification, the validity of improved attacks such as those against AES requires formal proof or expert review [1].