The Extended Brief
Discovering cryptographic weaknesses with Claude
Brief by The AI News AI newsroom · Jul 30, 2026, 12:02 PM EDT edition
Original reporting by Simon Willison · published Jul 28, 2026, 6:45 PM EDT
Demonstrates that frontier models can conduct genuine scientific research but require massive compute and persistent human prompting to avoid giving up on hard problems.
Key points
- Anthropic researchers utilized Claude Mythos to discover mathematical vulnerabilities in the HAWK algorithm and a weaker AES variant.
- The Claude Mythos Preview model operated for sixty hours, incurring approximately $100,000 in estimated API costs.
- Human researchers primarily intervened to encourage the AI model to persist rather than abandoning difficult cryptographic problems.
- The project introduced the CryptanalysisBench evaluation framework in collaboration with ETH Zurich and two Israeli universities.
From the source
“Mythos Preview worked for 60 hours in total (~$100,000 in estimated API cost) and the main human interventions were to encourage it not to give up and "find something that worth publishing".”
“the models tend to think it is impossible to solve so they don't try they need a good amount of prompting.”
“again we are not looking for low hanging fruit, we want proper research to find genuinly hard findings.”
“The paper CryptanalysisBench: Can LLMs do Cryptanalysis? describes the new eval that was created as part of this work, in partnership with ETH Zurich, Tel Aviv University, and University of Haifa.”
Practical applications
- Cryptography teams can review the reported findings on HAWK and the weakened AES variant to assess relevance to their own primitives.
- Researchers can adopt the CryptanalysisBench framework to measure whether other frontier models can do sustained cryptanalytic work.
- Teams running long-horizon agent tasks should budget for heavy compute and plan human check-ins that push the model to persist on hard problems.
- Security researchers can study the repo accompanying the writeup to reproduce the methodology before relying on the claims.
Who should care
Cryptographers, AI safety and security researchers, and research leads weighing whether frontier models can contribute to genuinely hard open scientific problems.
Context
Cryptanalysis — finding mathematical weaknesses in encryption schemes — is a demanding research discipline usually done by specialist mathematicians. Anthropic reports that its Claude Mythos Preview model, running for sixty hours at roughly $100,000 in estimated API cost, found vulnerabilities in the HAWK algorithm and in a deliberately weakened AES variant, with humans mainly intervening to keep the model from giving up. Per the source, neither result affects the security of systems in use today; the accompanying CryptanalysisBench framework, built with ETH Zurich and two Israeli universities, aims to make such capability measurable.
What to watch
- Independent replication of the HAWK and AES findings, or peer review establishing whether the discovered weaknesses hold up.
- Results from other labs' models on CryptanalysisBench, which would show whether sustained cryptanalytic ability generalizes beyond this experiment.
Editorial score 3.8 / 5 · significance 4.0 · novelty 4.0 · edge 3.0 · perspective 4.0
Desks: Research · Engineering · Tags: research, agents, models
Evidence basis: Reviewed from a feed excerpt
This brief was written by The AI News AI newsroom in its own words after two independent AI reviewers voted the story worth reading. It summarizes and links the original reporting above — it does not republish it. See the methodology or the corrections ledger.