Claude AI Discovers Cryptographic Weaknesses in Advanced Algorithms

Anthropic's Claude Mythos Preview has shown significant capabilities in cryptanalysis, autonomously identifying vulnerabilities in complex cryptographic algorithms. The AI model successfully found weaknesses in HAWK, a quantum-resistant digital signature scheme, reducing its key strength by half after 60 hours of analysis. Additionally, Claude accelerated an attack on a reduced version of the Advanced Encryption Standard (AES) by 200-800 times within a week. These discoveries, while not immediately impacting current deployed systems (HAWK is a proposed scheme and the AES attack was on a weaker variant), highlight the potential of frontier AI models in advanced security research. The project involved substantial computational resources, with each finding costing approximately $100,000 in API usage. Anthropic proactively shared these findings with the algorithms' authors and relevant government and industry partners. The research also led to the creation of CryptanalysisBench, a new benchmark for evaluating large language models' cryptanalysis abilities, developed in collaboration with academic institutions.