Details
- Anthropic reports that its Claude Mythos Preview model discovered weaknesses in two cryptographic primitives: the HAWK digital signature scheme and a reduced-round variant of the Advanced Encryption Standard (AES).
- HAWK is a post-quantum digital signature proposal designed to remain secure even against future quantum computers, and had previously withstood two years of expert review.
- Claude Mythos Preview found a new attack on HAWK within roughly 60 hours, cutting the effective key strength by about half, despite the scheme’s prior academic scrutiny.
- On the symmetric side, the model identified a way to accelerate an attack on a weakened version of AES by 200–800×, even though AES has been one of the most extensively analyzed encryption algorithms over the past decades.
- Anthropic says Mythos Preview carried out most of the cryptanalysis autonomously, with limited human guidance, and each of the two discoveries cost around $100,000 in API usage.
- The company disclosed the findings in advance to the original algorithm designers, relevant US government agencies, and industry partners to support responsible handling of the new weaknesses.
- Anthropic emphasizes that the results have no direct practical impact on current systems, since HAWK is not deployed and the AES attack targets only a reduced-strength construction, not the full standard.
- The work is framed as evidence that frontier AI models can perform expert-level cryptography research, with potential defensive uses in stress-testing cryptographic building blocks before real-world deployment.
- Anthropic has released technical papers detailing both attacks, along with model chain-of-thought logs for the AES result, to support independent verification and further study.
- In collaboration with researchers at ETH Zurich, Tel Aviv University, and the University of Haifa, Anthropic also introduced CryptanalysisBench, a benchmark designed to systematically evaluate and compare large language models’ capabilities in cryptanalysis.
- CryptanalysisBench aims to provide structured tasks and metrics for measuring how well LLMs can analyze and break cryptographic schemes, helping the research community track progress and risks over time.
Impact
Anthropic’s results signal that leading language models are beginning to meaningfully contribute to original cryptographic research, not just re-deriving known attacks. By demonstrating autonomous, costly but effective cryptanalysis on advanced schemes, Anthropic highlights both the promise of AI-assisted defensive testing and the need for policymakers and standards bodies to account for AI-accelerated discovery of vulnerabilities in future cryptographic designs.
