Discovering cryptographic weaknesses with Claude

You might assume AI will automatically tackle tough problems, but in reality, it behaves more like an intern who needs constant encouragement. Anthropic recently ran Claude Mythos on a cryptography task for 60 hours, spending roughly $100,000 in API costs. The result? It identified mathematical flaws in the HAWK protocol and a weakened version of AES. While the researchers explicitly noted these findings don't pose a real-world security threat, the real story isn't the outcome—it's the process.

The most revealing part is the shared prompt history. It shows human researchers repeatedly coaxing the model to keep working: when the AI hits a wall, it tends to conclude the problem is unsolvable and gives up. Humans had to constantly remind it, "We're looking for something worth publishing," and "Don't settle for low-hanging fruit—find genuinely hard breakthroughs." The prompts even preserve typos, making the exchange feel unusually authentic.

This points to a deeper shift: AI-driven research is moving from tool-assisted workflows to autonomous exploration, yet current models still lack true scientific resilience. They excel at pattern recognition and rapid validation but tend to retreat when faced with genuinely difficult challenges. The human role is no longer about writing code or running experiments—it's about project management and psychological coaching, setting direction, managing expectations, and preventing premature abandonment.

For everyday developers, this carries several implications. First, don't expect AI to automatically solve your hardest technical problems. It's more like a highly capable but easily distracted assistant that needs continuous context and motivation. Second, if you're doing research-oriented work, you can adopt this human-AI collaboration model: let the AI handle exhaustive search and verification, while you focus on steering direction and maintaining momentum. Finally, the CryptanalysisBench benchmark developed in partnership with top universities shows that academia is already building systematic ways to evaluate AI's capabilities in rigorous fields like cryptography.

Interestingly, Anthropic didn't hype "AI breaks AES." Instead, they openly acknowledged the findings pose no practical threat. That restraint actually builds more trust in their research quality. As models improve in complex reasoning, we'll likely see more AI research assistant cases, but the core bottleneck may shift from compute to keeping AI motivated when facing the unknown.

You might think AI is replacing scientists, but it's actually redefining what research looks like: moving from individual flashes of insight to reproducible, evaluable, and collaborative engineering processes. For every tech professional, this is a clear signal: future competitiveness won't be about whether you can code, but whether you can design effective AI workflows and make sound judgments at critical moments.