Anthropic’s Claude AI Discovers Critical Flaw in NIST Post-Quantum Cryptography Standard

AI Uncovers Hidden Weakness in HAWK Signature Scheme

Anthropic announced that its Claude Mythos Preview model has discovered a previously unknown attack on HAWK, a digital signature system currently advancing through the U.S. National Institute of Standards and Technology’s post-quantum cryptography competition. The vulnerability dramatically reduces the computational cost required to compromise the scheme’s smallest configuration—dropping the difficulty from 2^64 operations to 2^38, representing a reduction of approximately 67 million times.

HAWK functions as a mathematical system designed to authenticate transactions while remaining resistant to attacks from quantum computers, a capability that traditional cryptography lacks. The model identified an underlying symmetry in HAWK’s mathematical structure that human researchers had not previously discovered. While this finding does not immediately endanger deployed systems—HAWK has never entered production use—it illustrates the inherent complexity of designing cryptographic standards that can withstand both current and future computational threats.

Blockchain Block Space and Signature Efficiency at Risk

The implications carry particular weight for cryptocurrency networks, where cryptographic choices directly affect operational economics. The size of digital signatures correlates directly with block space consumption, which in turn influences transaction fees across blockchain networks. HAWK’s appeal as a post-quantum candidate rests primarily on its compact signature dimensions and rapid signing performance—characteristics that make it attractive for distributed ledgers seeking quantum-resistant alternatives to ECDSA, the signature algorithm currently used by Bitcoin.

Remedying the identified weakness would require approximately doubling HAWK’s key size, according to Anthropic. This correction substantially reduces the scheme’s efficiency advantage, potentially compromising its competitive standing among other post-quantum candidates. As blockchain projects evaluate quantum-resistant replacements for legacy signature systems, the security-efficiency tradeoff revealed by this discovery becomes a consequential factor in standards selection.

Responsible Disclosure and Expanded Cryptanalytic Capabilities

Anthropic disclosed the finding to HAWK’s designers, relevant U.S. government agencies, and industry stakeholders before public disclosure, maintaining coordination with NIST throughout. In addition to HAWK, the Claude model identified an accelerated attack against AES, the encryption cipher protecting HTTPS connections and encrypted data storage systems globally. The model developed a novel cryptanalysis method targeting a 7-round variant of AES that cryptographers had not improved upon since 2013, achieving 200 to 800 times faster performance than the previous record.

Each discovery required approximately $100,000 in computational costs, with Anthropic teams spending several hundred hours validating the AES research. These findings demonstrate an emerging capacity of advanced AI systems: autonomous identification of security vulnerabilities in core cryptographic algorithms, a capability historically restricted to specialized human researchers.

HAWK’s vulnerability demonstrates that post-quantum cryptography standards must survive AI-assisted cryptanalysis before standardization—a critical consideration for any blockchain transitioning away from ECDSA signatures.

Source: Anthropic, via Decrypt. Not financial advice.