We explore groundbreaking findings from Anthropic that highlight the Claude AI model’s potential to enhance security in the cryptocurrency domain.
Researchers at Anthropic utilized the Claude Mythos Preview model to uncover two forms of cryptanalytic assaults—one targeting the post-quantum HAWK digital signature scheme and another aimed at a reduced-round AES-128 variant. The organization underscored that these findings do not impact existing production systems and do not necessitate any adjustments to current software implementations.
Trending Topic: Bitcoin Treasury Hyperscale Liquidates 9% of Its BTC Despite Indefinite Hold Commitment
“These are noteworthy advancements in research, but they have no immediate implications for existing systems. Both findings demonstrate that cutting-edge AI models are adept at performing cryptography research at an expert level,” stated Anthropic.
Each of the two investigations incurred an estimated cost of around $100,000 in API credits. The process of exploring, developing, and validating the HAWK attack spanned approximately 60 hours, while confirming the AES outcome required several hundred research hours.
Contents
Insights from Claude Mythos on HAWK and AES: 60 Hours and $100,000 Expenses
HAWK is being considered for post-quantum digital signature standardization in the NIST competition. The Claude Mythos Preview unveiled an existing symmetry within HAWK’s mathematical framework that facilitated faster key recovery assaults.
- For HAWK-256, the anticipated cost for complete key recovery was reduced from 2^64 to 2^38 operations.
- For HAWK-512 and HAWK-1024, the attack remains nearly impractical.
Anthropic estimates that to preserve equivalent security levels, HAWK would need to nearly double its key sizes—compromising one of the scheme’s main benefits: compactness. The researchers clarified that this result does not extend to other NIST candidates and does not imply the failure of lattice-based post-quantum cryptography.
In another finding, Anthropic indicated that Claude Mythos Preview identified a vulnerability within 13 rounds of the Korean LEA block cipher standard, permitting key recovery in under an hour on a contemporary desktop computer. The full standard utilizes 24 rounds, indicating no immediate threat to existing systems.
Discover more: China’s Kimi K3 Surpasses Claude Fable 5 in Frontend Development — Asian AI Models Catch Up to US Leaders
Significance for Cryptocurrency: AI Boosts Cryptanalysis Without Compromising Security
Anthropic’s insights illustrate the capabilities of cutting-edge AI models in cryptographic analysis. The organization shared these findings with the developers of the algorithms, U.S. government entities, and industry collaborators beforehand. Regarding HAWK, Anthropic informed the scheme’s creators in June and aligned the announcement with NIST’s communication channels—adhering to ethical disclosure practices.
For the cryptocurrency sector, this signals that AI is becoming an essential instrument for evaluating cryptographic algorithms prior to their deployment. However, as AI capabilities expand, novel strategies will be necessary to handle the responsible disclosure of potentially significant vulnerabilities. Researchers emphasized that Anthropic’s discoveries represent research findings rather than practical instruments against today’s systems.
Explore further: Can Rogue AI Compromise Your Cryptocurrency? Understanding the OpenAI Scare and Its Impact on Investors
CryptanalysisBench: Paving the Path for Future AI Research
To assess the cryptanalytic capacities of language models, Anthropic, along with scholars from ETH▲$2,518.27 Zurich, Tel Aviv University, University of Haifa, and TU Berlin, launched the CryptanalysisBench benchmark. This benchmark encompasses 191 tasks across six categories of cryptographic primitives, primarily sourced from NIST competitions. Models are required not only to identify vulnerabilities but also to produce a functioning attack script that undergoes formal verification.
The authors reported that five advanced models tackled 65% to 86% of level-one challenges. Claude Mythos 5 achieved an accuracy of 85.7%, while the least effective AI assistant managed 65.3%. Performance was notably lower in complete schemes with no available practical attacks—no model surpassed 9%.
Discover more: What’s Next in Cryptocurrency After AI Tokens? Key Sectors to Monitor in H2 2026