Anthropic研究团队利用Claude Mythos Preview AI模型在密码学研究中取得重要发现。[1]该模型在60小时内改进了对后量子数字签名方案HAWK的攻击，将HAWK-256的破解成本从2^64降低至2^38，相当于将密钥强度减半。[1]同时，研究团队还发现了对7轮AES加密的新型攻击方法（Möbius Bridge），其速度相比现有方案提升了200至800倍。[1]
两项攻击研究的API调用成本均约为10万美元，其中AES攻击中模型输出了超过10亿个tokens，跨越三天完成。[1]虽然这些发现不会影响现有生产系统的安全，但充分展示了AI在密码分析领域的应用潜力。[1]Anthropic研究人员已与美国国家标准与技术研究所（NIST）、学术机构和政府合作伙伴进行了负责任的信息披露协调。[1]此外，Anthropic还与瑞士联邦理工学院、特拉维夫大学和海法大学合作开发了CryptanalysisBench基准，用于评估AI在密码学任务中的能力。[1]
Anthropic's research team has leveraged the Claude Mythos Preview AI model to uncover major weaknesses in cryptographic algorithms, demonstrating artificial intelligence's emerging capability in cryptanalysis.[1] Over a 60-hour period, the model substantially improved attacks against HAWK, a post-quantum digital signature scheme, reducing the key strength of HAWK-256 from 2^64 to 2^38.[1] Additionally, Claude Mythos Preview identified a novel attack method called Möbius Bridge against seven-round AES encryption, achieving speed improvements of 200 to 800 times over existing approaches.[1]
The research operations required significant computational resources, with each attack—the HAWK assault and the AES breakthrough—consuming approximately $100,000 in API costs.[1] The AES attack alone generated over one billion tokens across a three-day computation span.[1] While these discoveries do not compromise existing production systems currently in use, they illustrate AI's potential as a tool for advancing cryptographic research and identifying previously unknown attack vectors.[1] Anthropic coordinated responsible disclosure of these findings with the U.S. National Institute of Standards and Technology (NIST), academic institutions, and government partners.[1] The research was conducted in collaboration with ETH Zurich, Tel Aviv University, and the University of Haifa, which together helped develop the CryptanalysisBench benchmark.[1]