Matthew Green Analyzes Claude Mythos Cryptanalysis Research

AI models paired with agentic harnesses are evolving from code assistants into active scientific research partners, though theoretical cryptanalysis breakthroughs must be distinguished from immediate real-world security vulnerabilities.

• Anthropic's Claude Mythos model discovered a theoretical key recovery attack against HAWK that halves its bit-security margin, making the candidate scheme significantly less viable for future standardization.

• The attack on 7-round AES represents an incremental optimization of a decade-old theoretical result and poses no threat to the full, standard AES encryption used today.

• Anthropic achieved these discoveries through automated agentic loops pushing LLMs to systematically explore mathematical hypotheses rather than relying purely on manual expert prompting.

• The milestone highlights AI's potential to accelerate complex mathematical research while raising important questions about the future role of human cryptographers.