Anthropic's Claude Mythos Preview found a critical weakness in HAWK, a post-quantum digital signature scheme under NIST review, and developed a faster attack on reduced-round AES, raising new questions about cryptographic security in the quantum era
An artificial intelligence model has identified a significant vulnerability in HAWK, a post-quantum digital signature scheme currently being evaluated by the U.S. National Institute of Standards and Technology (NIST) for future cryptographic standards. According to Anthropic, its Claude Mythos Preview model discovered the flaw in HAWK in just 60 hours-despite the scheme having previously passed two years of expert human review. The same AI system also developed a more efficient attack on a reduced-round version of the Advanced Encryption Standard (AES), a widely used symmetric cipher.
HAWK is one of several remaining candidates in NIST's ongoing effort to select cryptographic algorithms that could withstand attacks from large-scale quantum computers. Quantum computing poses a threat to most current encryption methods, making the search for quantum-resistant alternatives a top priority for governments and industry. The vulnerability found in HAWK stems from a mathematical property called a nontrivial automorphism-a symmetry in the lattice structure that underpins the scheme's security. Prior research had suggested that such a symmetry could enable an attack, but it was not previously known to exist in HAWK's design. Exploiting this property effectively halves the scheme's key strength, meaning that restoring its original security level would require doubling key sizes. This change would undermine HAWK's competitiveness as a practical post-quantum solution.
Anthropic reports that the discovery process involved one researcher with a theoretical computer science background but no specialized expertise in lattice-based cryptography. The Claude Mythos Preview model operated within a multi-agent framework using Claude Code, with access to Python, Sage, and published cryptographic literature. Human involvement was limited to project management, and the process cost approximately $100,000 in API usage over 60 hours. The company shared its findings with HAWK's authors in June 2026 and coordinated public disclosure with NIST and academic partners.
AI-Driven Cryptanalysis
In a separate effort, Anthropic's model was tasked with improving attacks on a reduced version of AES, specifically a seven-round variant (the standard uses ten rounds). Initially, the model declined to pursue the problem, stating that surpassing existing cryptanalysis was unlikely. After a single prompt to seek new approaches, it developed a technique called the Möbius Bridge, which removes one of the guesses required in a meet-in-the-middle attack. This innovation made the attack between 200 and 800 times faster than the previous best-known method, depending on the measurement criteria. Two Anthropic researchers, neither with deep cryptography backgrounds, spent several hundred hours validating the result.
It is important to note that neither the HAWK nor the AES findings affect any cryptographic systems currently in production. HAWK has not been deployed, and the AES result applies only to a reduced-round version, not the full standard used in practice. Still, the discoveries highlight the potential for advanced AI models to accelerate cryptanalysis and challenge assumptions about the security of cryptographic candidates, even those that have undergone extensive expert review.
Industry and Research Response
Anthropic coordinated its disclosure with researchers at ETH Zurich, Tel Aviv University, and the University of Haifa, as well as with U.S. government and industry stakeholders. Alongside the technical findings, the company released CryptanalysisBench, a benchmark suite designed to track AI performance on cryptanalytic tasks over time. Anthropic stated that work on additional ciphers is ongoing, with further results expected after validation.
The rapid identification of a critical weakness in a NIST candidate by an AI model raises questions about the future of cryptographic review and the role of automated systems in protocol evaluation. As quantum computing advances, the need for robust, thoroughly vetted post-quantum cryptography becomes more urgent. The HAWK incident demonstrates that even schemes considered strong by human experts may harbor subtle vulnerabilities detectable by AI-driven analysis.
According to NIST's official timeline, the agency expects to finalize its initial set of post-quantum cryptographic standards by 2027. The HAWK discovery may prompt additional scrutiny of remaining candidates and influence the selection process. For now, the incident serves as a reminder that cryptographic security is a moving target-and that both human and machine intelligence will play a role in shaping the next generation of digital protection.
As of June 2026, NIST's post-quantum cryptography project is in its final evaluation phase, with several digital signature and key encapsulation schemes under review. HAWK was among the few remaining digital signature candidates before the reported vulnerability. The AES cipher, in its standard 10-round form, remains widely used in global financial, government, and blockchain applications, with no known practical attacks against the full version. The cost and speed of the AI-driven cryptanalysis-$100,000 and 60 hours for HAWK, three days and roughly 1 billion output tokens for the AES attack-underscore the growing capabilities of advanced language models in technical domains.
Cryptographic Symmetries and Security Trade-Offs
Modern cryptography relies on mathematical structures that are assumed to be hard to break, such as lattices, elliptic curves, or large prime factorizations. When a cryptographic scheme contains hidden symmetries or structural weaknesses, attackers may exploit these to reduce the effective security level, sometimes dramatically. In the case of HAWK, the presence of a nontrivial automorphism in its lattice structure allowed the AI model to devise an attack that cut the scheme's key strength in half. This kind of vulnerability is difficult to detect through manual review alone, especially as cryptographic designs become more complex. The incident highlights the importance of diverse review methods-including automated analysis-when evaluating protocols intended to secure digital assets against future quantum threats.
