Anthropic's AI model, Claude Mythos Preview, recently successfully identified mathematical vulnerabilities in the encryption algorithms that underpin global digital security, demonstrating the autonomous reasoning and attack capabilities of large models in complex cryptography fields.

In research targeting the post-quantum signature scheme candidate algorithm HAWK from the U.S. National Institute of Standards and Technology (NIST), Mythos developed an improved attack by exploiting previously unnoticed symmetries in "mathematical lattices" within a multi-agent semi-autonomous system. This breakthrough solved a problem that had been reviewed by human experts for over two years, all within 60 hours and with a $100,000 API cost.

Anthropic, Claude

Additionally, regarding the most widely used symmetric encryption standard in the world—simplified AES-128 (7-round modified version)—Mythos independently developed a new fingerprint identification attack method called "Mobius Bridge" with only three intervention prompts and consuming 1 billion tokens (approximately $100,000 cost). This method increased attack efficiency by 200 to 800 times.

Although Anthropic emphasized that these breakthroughs do not affect currently running systems, and that Mythos Preview is still not open to the public, the company has completed vulnerability coordination disclosure with the U.S. government, industry partners, and HAWK developers. It has also collaborated with institutions such as ETH Zurich to launch the CryptanalysisBench benchmark test. This development indicates that AI is posing unprecedented challenges to the core assumptions underlying internet security.