Artificial intelligence is reshaping the landscape of cryptographic security, uncovering previously unseen vulnerabilities in widely used encryption schemes. The latest findings from Anthropic’s Frontier Red Team demonstrate how advanced large language models can identify and exploit weaknesses in cryptographic frameworks, pushing the boundaries of what was previously considered secure.

AI’s Role in Cryptographic Vulnerability Discovery

The Frontier Red Team at Anthropic has published a detailed analysis of two critical cryptographic algorithms: HAWK, a post-quantum digital signature scheme, and a reduced-round variant of AES. Their work highlights the growing influence of AI in cryptographic research, where machine learning models are becoming indispensable tools for identifying potential flaws before they can be exploited in real-world scenarios.

Anthropic’s research team utilized a pre-release version of their Mythos language model to systematically analyze HAWK and AES. The AI-driven approach allowed them to discover vulnerabilities that traditional cryptographic analysis had missed, demonstrating how AI can accelerate the discovery of security flaws while also raising new questions about the robustness of widely deployed encryption standards.

Breaking Down the HAWK Attack

HAWK was a leading candidate in the NIST Post-Quantum Cryptography Standardization Process, designed to withstand attacks from quantum computers. Despite passing two rounds of rigorous human expert evaluations, HAWK was recently compromised by Anthropic’s AI-driven analysis. The breakthrough came when the team identified a geometric symmetry in the algorithm’s underlying lattice structure, which reduced the complexity of key recovery attacks.

By exploiting this symmetry, researchers were able to cut the computational effort required to extract private keys in half. For example, on the HAWK-256 variant, the attack reduced the key recovery complexity from 2⁶⁴ to 2³⁸ operations. This breakthrough was achieved in just 3 hours and 42 minutes using a standard 96-core server, showcasing the efficiency of AI-assisted cryptanalysis.

The implications of this discovery are significant. The HAWK team has since withdrawn the algorithm from the NIST selection process, effectively rendering it obsolete. The attack demonstrates that even algorithms vetted through extensive human review processes can harbor hidden vulnerabilities that AI-driven analysis can uncover.

AES in the Crosshairs: A Theoretical Breakthrough

The second major finding involves a reduced-round variant of AES-128, a widely used symmetric encryption standard. While the full AES algorithm remains secure, Anthropic’s AI identified a critical flaw in a simplified version of the cipher, restricted to just 7 rounds instead of the standard 10.

The breakthrough came through an innovative technique called the Möbius Bridge, an invariant fingerprinting method integrated into a meet-in-the-middle attack. This approach eliminated an exhaustive search phase previously requiring 256 iterations, accelerating the attack by a factor of 200 to 800 compared to prior benchmarks from 2013.

Importantly, this vulnerability remains strictly confined to the academic realm. The additional three rounds in the full AES-128 standard render such attacks impractical, ensuring that real-world applications remain unaffected. The findings underscore the importance of testing simplified versions of encryption algorithms to anticipate potential weaknesses before they can be exploited in full-scale implementations.

AI’s Growing Influence on Cryptographic Research

The Anthropic team’s work highlights a paradigm shift in cryptographic research. Traditionally, cryptanalysis relied on human expertise, intuition, and painstaking manual analysis. While AI-driven approaches are not yet replacing human cryptographers, they are becoming an increasingly vital complement to traditional methods.

The research process itself was revealing. The AI initially refused to perform the analysis on the reduced-round AES variant, asserting that the algorithm was too robust for improvement. After refining the prompts and scaffolding, the model eventually generated a novel attack vector, producing roughly one billion tokens over several days. The computational cost of such an operation would have exceeded $100,000 if performed using commercial AI services.

This experiment underscores both the potential and limitations of AI in cryptography. While AI can uncover vulnerabilities at an unprecedented pace, human oversight remains essential to validate findings, ensuring mathematical rigor and eliminating the risk of hallucinated results. The Anthropic team spent over a month verifying the AI-generated attack vectors, conducting formal checks on the mathematical proofs, algebraic transformations, and complexity reductions proposed by the model.

The Future of Cryptographic Security

These discoveries come at a time when the cybersecurity community is already grappling with an overwhelming volume of vulnerabilities. AI-driven tools are accelerating the pace of discovery, but they are also creating new challenges for triage, verification, and remediation. The industry is shifting from a focus on finding flaws to prioritizing the validation and implementation of fixes.

The Anthropic research serves as a reminder of the need for ongoing vigilance in cryptographic security. While the immediate impact of these findings is limited, they signal a broader trend: AI is becoming an integral part of cryptographic research, and its role will only grow in the coming years. As AI models become more sophisticated, they will likely uncover new classes of vulnerabilities, necessitating adaptive strategies to maintain cryptographic robustness in an evolving threat landscape.

The findings also highlight the importance of collaboration between AI researchers and cryptographers. By combining the strengths of both disciplines, the cybersecurity community can better anticipate and mitigate emerging threats.