Anthropic's Claude Mythos Preview has revealed critical vulnerabilities in cryptographic algorithms, including HAWK and AES, raising concerns over quantum resistance.
Summary
- Anthropic's Claude Mythos Preview model has identified a new method to attack HAWK, decreasing the effort to breach its smallest key from 2^64 to 2^38 operations.
- The model also achieved an 800-fold acceleration in attacking a 7-round version of AES, surpassing previous cryptographic records.
- These findings cost approximately $100,000 in API usage, with researchers dedicating hundreds of hours to validate the results.
Anthropic has announced that its upcoming AI model, Claude Mythos Preview, has uncovered significant vulnerabilities in cryptographic algorithms, focusing on HAWK, a candidate for a federal digital signature standard in the U.S.
HAWK aims to secure digital signatures against quantum threats and has progressed to the third round of the National Institute of Standards and Technology (NIST) post-quantum signature competition, where it is the only lattice-based candidate.
The model's analysis revealed a previously unnoticed mathematical symmetry in HAWK, allowing the recovery of a secret key with only 2^38 operations—about 67 million times easier than before. To address this flaw, HAWK would need to double its key sizes, a change that could undermine its advantages as a post-quantum cryptography candidate.
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
This adjustment is crucial for blockchain applications, as signature size impacts block space and transaction fees. HAWK's original appeal lay in its compact key size and efficient signing speed, attributes that would be compromised by the necessary adjustments.
For cryptocurrency users, there is no immediate risk since HAWK has not yet been implemented, and Bitcoin continues to use the ECDSA signature scheme that HAWK aims to replace.
Before publicizing its findings, Anthropic shared its discoveries with the algorithm's authors and relevant U.S. government and industry stakeholders, ensuring coordination with NIST regarding the HAWK findings.
Breakthrough in AES Cryptanalysis
The second vulnerability identified pertains to the AES encryption standard, which secures various digital communications and systems. While the full AES-128 standard involves 10 rounds of encryption, Claude focused on a less-studied 7-round variant.
Researchers set stringent conditions for the testing, restricting the model from using established AES cryptanalysis methods and pushing it to devise a novel approach. Despite the challenges, Claude succeeded in creating a technique known as a Möbius Bridge, which reduced the number of key bytes an attacker needs to guess from nine to eight after extensive analysis.
Additionally, Claude demonstrated the ability to break 13 rounds of LEA, a national encryption standard from Korea, in under an hour, outperforming previous best efforts that required an extensive number of plaintext pairs. However, the standard implementation of LEA remains secure, as it runs 24 rounds.
Longer Verification Process
The HAWK research paper highlights that most mathematical discoveries were facilitated by AI, with human contributions focused on oversight and validation. Although Claude identified the AES vulnerability quickly, researchers spent significant time ensuring its accuracy and reliability.
Anthropic's previous assessments have shown that AI models can uncover numerous vulnerabilities at a pace that may outstrip traditional human verification processes, raising concerns about potential bottlenecks in cybersecurity.
The company has also developed CryptanalysisBench, a platform that includes various cipher-breaking tasks derived from NIST competitions, to measure the effectiveness of their models in real-world scenarios.
