Anthropic's Claude finds flaws in HAWK and AES encryption

What happened

Anthropic researchers used Claude Mythos, a reasoning model, to discover mathematical weaknesses in the HAWK encryption algorithm and a reduced version of AES (a widely used encryption standard). The model worked for 60 hours in total, with an estimated API cost of ~$100,000, guided by researchers through prompts that emphasized finding novel, publishable results rather than obvious vulnerabilities.

Why it matters

The research demonstrates that AI reasoning models can assist in cryptographic analysis—a task typically requiring expert human mathematicians. The article notes that "neither of these results has a practical impact on today's computer systems", but the work shows how large language models can be prompted to persist through difficult problems when instructed not to give up and to seek genuinely hard findings rather than easy targets.

What to watch

The shared prompts reveal the iterative nature of guiding Claude Mythos through research: researchers had to repeatedly redirect the model away from giving up, clarify that the goal was novel attacks (not minor tweaks to existing methods), and stress the need for publishable-quality findings. The work is documented in a public repository, allowing others to study how AI can be steered toward rigorous research outcomes.