Anthropic's Claude AI Weakens Post-Quantum HAWK Scheme, Finds AES Attack Vector

Anthropic researchers deployed their Claude Mythos Preview AI model to identify improved attacks on cryptographic algorithms, marking a shift from automated bug-hunting to high-level mathematical cryptanalysis.

The AI weakened HAWK, a post-quantum digital signature scheme based on the lattice isomorphism problem, and discovered a new attack method for round-reduced AES. The findings do not currently affect any production systems.

Targeting Post-Quantum Cryptography

HAWK is a candidate in the ongoing standardization process for cryptographic algorithms designed to withstand attacks from quantum computers. The research demonstrates that AI models can now engage with the mathematical foundations underlying these schemes, rather than simply identifying implementation errors in code.

Anthropic's research publication states the findings significantly weaken HAWK's security margins. The company did not disclose full technical details of the attack methodology in publicly available materials.

AES Attack Vector Discovered

The Claude model identified a new way to attack round-reduced AES. Round-reduced variants are simplified versions of the full cipher used primarily in academic research and cryptanalysis competitions, not in production deployments.

The full AES algorithm, which secures HTTPS connections and encrypted disk storage, remains unaffected. The attack applies only to reduced-round versions used for testing cryptographic resilience.

No Immediate Production Impact

Anthropic states these findings do not currently affect any production systems. HAWK has not been deployed at scale, and the AES attack targets reduced-round variants rather than the full cipher used in real-world applications.

The research follows earlier demonstrations of Claude Mythos Preview's ability to autonomously find and exploit software vulnerabilities. This project represents an escalation in capability, moving from code-level bugs to the mathematical primitives underpinning digital security.

Implications for Cryptographic Security

AI-assisted cryptanalysis could accelerate the erosion of security margins for existing encryption standards. As models improve, the timeline for maintaining confidence in current cryptographic assumptions may need to compress.

Security researchers have long anticipated machine learning would enter the cryptanalysis domain. This research provides concrete evidence that the transition from theoretical possibility to practical capability is underway.

The findings underscore the importance of proactive approaches to post-quantum cryptography and continuous evaluation of symmetric cipher security assumptions. Organizations relying on long-term cryptographic guarantees should monitor these developments as part of their security planning.