Anthropic's Claude AI model has demonstrated novel cryptanalytic capabilities by deriving a complete key-recovery attack against HAWK-256, a post-quantum lattice-based signature scheme, and accelerating attacks on reduced-round AES-128 by 200 to 800 times.

The HAWK attack exploits a previously unknown symmetry property in the lattice structure underlying the signature scheme. Anthropic's implementation achieves full key recovery in approximately three hours and 42 minutes on a 96-core server, representing a practical break of the algorithm's security claims.

The AES findings target only seven-round variants of AES-128, not the production 14-round standard. This remains a theoretical rather than practical threat to deployed systems. However, the speedup magnitude demonstrates AI's growing capability to uncover attack vectors humans may overlook.

These results emerged from Claude Mythos Preview, Anthropic's latest model iteration. The research underscores both the potential and the risk of deploying advanced AI systems in cryptanalysis. On the defensive side, AI can identify weaknesses in emerging post-quantum algorithms before they see widespread adoption. On the offensive side, the findings highlight how automated reasoning can discover novel mathematical properties with limited human guidance.

HAWK-256 specifically aimed to provide post-quantum security. The scheme entered academic consideration as organizations prepare cryptographic transitions away from RSA and elliptic curve methods vulnerable to quantum computing. This vulnerability discovery suggests security reviews for post-quantum candidates must now account for AI-assisted cryptanalysis.

The practical impact depends on context. HAWK-256 exists primarily in academic settings rather than production infrastructure, making the discovery valuable for algorithm refinement before standardization. The AES-128 attack on seven-round variants poses minimal risk to real systems, which implement significantly more rounds.

Anthropic released implementation details, allowing independent