Researchers at Anthropic, using their Claude Mythos Preview model, have identified significant vulnerabilities in cryptographic algorithms. The first discovery weakens HAWK, a digital signature scheme designed for a post-quantum era. Despite HAWK undergoing two rounds of expert review, Mythos reduced its effective key strength by half in just 60 hours. \n\nThe second breakthrough involves a new attack on a reduced version of AES, the most widely used symmetric cipher.

Mythos improved the speed of previous attacks by 200-800×, eliminating one necessary guess for an attacker. \n\nWhile these advancements do not currently impact production systems, they highlight the potential of advanced AI models to stress-test cryptographic algorithms. Anthropic has partnered with universities to create Cryptanalysis Bench, a benchmark for evaluating LLMs in this field, ensuring more robust digital security.