Anthropic's AI Model Cracks a Post-Quantum Algorithm That Survived Years of NIST Review

A post-quantum cryptographic algorithm that had spent years under rigorous evaluation has been withdrawn from contention as a potential US standard after Anthropic's AI security model, Mythos, helped identify a previously unknown vulnerability. The algorithm in question, HAWK, is a digital signature scheme built to resist attacks from future quantum computers. It had successfully navigated two full rounds of testing by the National Institute of Standards and Technology before Mythos found the flaw during a third evaluation round — the very stage designed to surface exactly this kind of critical weakness.

The discovery came from an Anthropic researcher with no specialized cryptography background who, working semi-autonomously alongside Mythos, prompted the model through an extensive literature review, mathematical reasoning sessions, and computational experiments. The process cost roughly $100,000 in compute time and around 60 hours of work. Mythos deployed two independent agents during the process: one initially dismissed a candidate attack method as unworkable, while the second found a viable path forward. The agents eventually converged and jointly confirmed the attack was effective, after which Mythos built a verification pipeline to validate the result.

The core of HAWK's security rests on the Lattice Isomorphism Problem, a mathematical challenge considered resistant to quantum computing attacks. The best-known classical attack exploits automorphism symmetries, and Mythos produced a previously undiscovered method for identifying those symmetries that effectively cut HAWK's key strength in half. While doubling the key size could patch the vulnerability in theory, doing so adds enough computational overhead to make HAWK less attractive than competing post-quantum signing algorithms like ML-DSA and FN-DSA. Following Anthropic's Monday announcement, HAWK's developer formally withdrew the algorithm on Tuesday.

Sopie Schmieg, a Google engineer specializing in post-quantum cryptography, noted that weaknesses in HAWK had long been suspected but not yet demonstrated with this level of impact. Johns Hopkins cryptography professor Matthew Green highlighted a particularly telling characteristic of the finding: the attack did not rely on any new mathematics, but instead stitched together several existing, well-known tools that no one had previously thought to combine. That kind of synthesis, Green suggested, is precisely where AI systems excel and human researchers can fall short.

Anthropic also reported a secondary finding against a weakened version of AES, where Mythos improved the best-known meet-in-the-middle attack by incorporating a Möbius Bridge fingerprinting technique, reducing required plaintext inputs from roughly 2^105 to 2^89. The company was careful to frame both results with important caveats: neither breaks any cryptosystem in active use today, both were tested against deliberately weakened challenge instances rather than production implementations, and the AES improvement remains computationally infeasible outside laboratory conditions. Despite those limitations, Anthropic argued that the findings mark a meaningful shift in the pace and nature of AI-assisted cryptanalysis.