Summary & Key Takeaways ​

Anthropic's Claude Mythos identified mathematical flaws in HAWK and a weaker AES version. The research involved 60 hours of Mythos Preview, costing an estimated $100,000 in API usage. Key human intervention focused on encouraging the model not to give up and to find publishable results. The paper "CryptanalysisBench: Can LLMs do Cryptanalysis?" describes a new evaluation framework. Researchers shared the specific, often misspelled, prompts used to guide Claude. The findings currently have no practical impact on existing computer systems.

Our Commentary ​

This is fascinating. We've been wondering about LLMs' ability to do genuine research. The prompts are gold; they show how much coaxing these models still need. It's a bit unsettling to think about the cost, but the potential is clear. I'm genuinely curious how this evolves.