Anthropic researchers using the Claude Mythos model have discovered mathematical vulnerabilities in cryptographic algorithms, marking a shift in AI from finding implementation errors to discovering fundamental mathematical weaknesses.
What Happened
The Claude Mythos model identified a lattice symmetry in the HAWK post-quantum digital signature scheme, reducing the effective key length of HAWK-256 from 2^64 to 2^38 operations. Additionally, a "Möbius Bridge" method was developed to attack a 7-round version of AES-128, allowing for an attack acceleration of 200–800 times by exploiting the algebraic structure of the S-box.
Context
This discovery demonstrates the ability of LLMs to perform high-level mathematical analysis and cryptanalysis, moving beyond simple software bug hunting in code.
Why It Matters for the Industry
These findings may necessitate an urgent review of post-quantum cryptography (PQC) standards being developed by NIST and changes to key sizes to maintain security. It also stimulates growing demand for specialized AI agents for automated security auditing and Red Teaming.
Why It Matters for Users
For users, this means AI is becoming a serious tool in cryptanalysis. While core standards, such as full AES, remain secure, the automation of vulnerability discovery could create a "verification crisis," where human experts cannot keep pace with verifying and fixing new AI-driven findings.
What Is Not Yet Known / Limitations
There is a divergence in the assessment of consequences: while product developers see this as an opportunity to create new verification tools, system architects are focusing on the risks and the need for protection against automated attacks.
Sources
- Anthropic Research
- Cryptanalysis of 7-Round AES via the Algebraic Structure of its S-box (Möbius Bridge)
Author
Look at AI, Editorial Staff