We break down new research from Anthropic showing how the Claude AI model can be used to strengthen security in the crypto sector.
Anthropic researchers used the Claude Mythos Preview model to discover two cryptanalytic attacks–one on the post-quantum HAWK digital signature scheme and one on a reduced-round AES-128 variant. The company emphasized that the results don’t affect production systems and don’t require changes to deployed software.
Hot topic:<a href="https://cryptoz7.com/a-tactical-edge-for-bitcoin-amplify-premieres-bnav/” title=”A Tactical Edge for Bitcoin: Amplify Premieres BNAV”>Bitcoin Treasury Hyperscale Sells 9% of Its BTC Despite Indefinite Hold Pledge
“These are significant research advances, but they have no practical impact on today’s systems. Both results show that frontier AI models are capable of conducting cryptography research at an expert level,” Anthropic said.
Each of the two studies cost roughly $100,000 in API credits. The search, development, and verification of the HAWK attack took about 60 hours; confirming the AES result took several hundred hours of researcher time.
New Anthropic research: Discovering cryptographic weaknesses with Claude.
Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms—the mathematical methods that are used to keep data private.
Read more: https://t.co/TYKLjb3Q7V
— Anthropic (@AnthropicAI) July 28, 2026
- What Claude Mythos Found in HAWK and AES: 60 Hours and $100,000 in Costs
- Why This Matters for Crypto: AI Accelerates Cryptanalysis but Doesn’t Break Systems
- CryptanalysisBench and the Future of AI Research
What Claude Mythos Found in HAWK and AES: 60 Hours and $100,000 in Costs
HAWK is a candidate for post-quantum digital signature standardization under NIST’s competition. Claude Mythos Preview discovered a previously unused symmetry in HAWK’s mathematical structure that accelerated key recovery attacks.
- For HAWK-256, the expected cost of full key recovery dropped from 2^64 to 2^38 operations.
- For HAWK-512 and HAWK-1024, the attack remains practically infeasible.
Anthropic estimates that to maintain the same security level, HAWK would need to roughly double its key sizes–weakening one of the scheme’s advantages: compactness. The researchers stressed that the result doesn’t transfer to other NIST candidates and doesn’t mean lattice-based post-quantum cryptography is broken.
Separately, Anthropic said Claude Mythos Preview found an attack on 13 rounds of the Korean LEA block cipher standard that could recover the key in under an hour on a modern desktop computer. The full version uses 24 rounds, so this result also poses no immediate risk to current systems.
Why This Matters for Crypto: AI Accelerates Cryptanalysis but Doesn’t Break Systems
Anthropic’s results show the potential of frontier AI models for cryptography analysis. The company disclosed the findings to the algorithm authors, US government agencies, and industry partners in advance. For HAWK, Anthropic notified the scheme’s creators in June and coordinated the release with NIST’s mailing list–following responsible disclosure practices.
For the crypto industry, it’s a signal that AI is becoming a key tool for testing cryptographic algorithms before they’re deployed. But as model capabilities grow, new approaches to responsible disclosure of potentially critical vulnerabilities will be needed. The researchers stressed that Anthropic’s findings are research results, not practical tools against modern systems.
CryptanalysisBench and the Future of AI Research
To evaluate language models’ cryptanalytic capabilities, Anthropic, together with researchers from ETH▲$1,761.17 Zurich, Tel Aviv University, University of Haifa, and TU Berlin, introduced the CryptanalysisBench benchmark. It includes 191 tasks across six families of cryptographic primitives, mainly from NIST competitions. Models aren’t just asked to explain vulnerabilities–they must submit a working attack script that passes formal verification.
According to the authors, five tested frontier models solved 65% to 86% of level-one tasks. Claude Mythos 5 scored 85.7%, while the weakest tested AI assistant scored 65.3%. On full schemes with no published practical attacks, results were much lower–no model exceeded 9%.
Learn more:What’s the Next Big Crypto Narrative After AI Tokens? Top Sectors to Watch in H2 2026
