Claude Identifies Flaws in HAWK and Simplified AES, Proving Cryptographic Research Capability

Anthropic's Claude autonomously identified weaknesses in post-quantum HAWK and a simplified AES variant. While it accelerates research, it cannot breach real-world systems or replace experts.

Woofun AI reports that Anthropic enabled Claude Mythos Preview to autonomously explore research papers, formulate hypotheses, write code, and run experiments to identify mathematical weaknesses in cryptographic algorithms. The model introduced two new attacks: one targeting the yet-to-be-deployed post-quantum signature scheme HAWK and another targeting a simplified version of AES. The former took about 60 hours to complete, while the core idea of the latter was mainly proposed by the model.

Claude breached HAWK-256, a testing variant designed to withstand future quantum computers, lowering security estimates for larger parameters. Regarding AES, widely used by websites and banks, the full standard remains uncompromised; experiments targeted a 7-round version with three rounds of protection removed, requiring approximately 2^105 sets of specialized inputs. The claimed speedup only accelerates a theoretically unfeasible attack. Claude can expedite research but cannot replace experts or directly breach real-world systems.