AI model Mythos uncovers weakness in post-quantum cryptography candidate HAWK

Anthropic reported that its Mythos AI security model identified a previously unknown weakness in HAWK, a lattice-based digital signature scheme designed to resist quantum computer attacks and under review as a candidate for post-quantum cryptographic (PQC) standards. The flaw was discovered after about 60 hours of work and approximately $100,000 in compute costs, according to the company.

The attack leveraged Mythos to exploit the Lattice Isomorphism Problem, a mathematical foundation believed to be resistant to quantum attacks. By identifying a new method to find automorphism symmetries within this problem, Mythos effectively halved the effective key strength of HAWK. While the weakness can be mitigated by doubling the key size, the added computational overhead makes HAWK less competitive compared to other PQC signing algorithms such as ML-DSA and FN-DSA.

The developer of HAWK withdrew the algorithm from consideration following the discovery, despite HAWK having previously withstood two rounds of NIST testing. The flaw was uncovered during a third round of testing designed to catch precisely such weaknesses, highlighting the evolving nature of cryptanalysis and the potential for AI to uncover subtle vulnerabilities.

Mythos operated in a semi-autonomous agentic workflow, conducting an extensive literature review, performing mathematical reasoning, and running computational experiments. The model implemented an end-to-end verification pipeline to validate the attack’s correctness. Notably, two separate agents initially disagreed on the viability of the method before collaborating to confirm the improved attack’s effectiveness.

Experts noted that the attack does not break real-world cryptosystems but represents a meaningful advance in AI-assisted cryptanalysis. Matthew Green, a Johns Hopkins cryptography professor, emphasized that the attack combines existing methods in a novel way, making it particularly suited for AI-driven discovery. Sophie Schmieg, a Google expert in PQC, stated that the findings effectively ended HAWK’s viability as a competitive candidate, describing the result as making "HAWK dead."