Anthropic's Claude Mythos AI Finds Weaknesses in HAWK Post‑Quantum Signature and Reduced‑Round AES
Anthropic’s Claude Mythos Preview model autonomously discovered a structural weakness in the HAWK‑256 lattice‑based post‑quantum digital signature scheme. The flaw reduces the scheme’s effective security from an estimated 2⁶⁴ operations to about 2³⁸, allowing a full key‑recovery attack in roughly 3 hours 42 minutes on a 96‑core server. The AI‑driven effort took about 60 hours of computation and cost around $100 000 in API usage. In a separate result, Mythos accelerated a meet‑in‑the‑middle attack on a 7‑round reduced version of AES‑128 by 200‑ to 800‑fold, though the attack still requires about 2¹⁰⁵ chosen plaintexts and remains impractical for real‑world systems. Anthropic confirmed that neither finding threatens any production cryptographic software. The HAWK weakness prompted the scheme’s developers to withdraw it from the National Institute of Standards and Technology’s (NIST) post‑quantum standardization process. The discoveries were publicly disclosed on 28 July 2026 and shared with NIST and industry partners.
Entities: AES‑128 · AES‑128 (7‑round variant) · AES‑128 (reduced‑round version) · Advanced Encryption Standard (AES) · Anthropic · Claude Mythos Preview · HAWK · HAWK (post‑quantum signature scheme) · HAWK‑256 · National Institute of Standards and Technology · National Institute of Standards and Technology (NIST)
Claims
What the coverage asserts, and how well corroborated each claim is across sources.
- [● 5 SOURCES] HAWK is a third‑round candidate in NIST’s post‑quantum cryptography standardization process and is not yet deployed. (NIST / Anthropic)
- [● 5 SOURCES] The HAWK attack was developed in about 60 hours of AI work costing roughly $100,000 in API usage. (All articles)
- [● 5 SOURCES] Claude Mythos Preview improved a meet‑in‑the‑middle attack on a seven‑round reduced AES‑128 cipher, making it 200‑800 times faster than previous methods. (All articles)
- [● 5 SOURCES] The HAWK attack remains exponential and impractical for larger key sizes, but would require doubling key sizes to restore security. (All articles)
- [● 5 SOURCES] Claude Mythos Preview discovered a structural weakness in the HAWK post‑quantum digital signature scheme that reduces its effective security from 2^64 to 2^38 operations. (All articles)
- [● 5 SOURCES] The findings illustrate that AI models can identify mathematical flaws in cryptographic algorithms that have evaded years of human expert review. (All articles)
- [● 5 SOURCES] The AES‑128 attack targets a 7‑round version, requires about 2^105 chosen plaintexts, and is considered completely impractical for real‑world use. (All articles)
- [● 5 SOURCES] Neither of the discovered attacks affect production cryptographic systems or full‑AES‑128 implementations. (All articles)
- [● 5 SOURCES] The attack on HAWK remains exponential and is not practical against larger parameter sets. (Anthropic researchers)
- [● 5 SOURCES] Claude Mythos Preview improved a meet‑in‑the‑middle attack on a seven‑round reduced version of AES‑128, making it 200–800 times faster than previous methods. (Anthropic)
- [● 5 SOURCES] The HAWK attack required about 60 hours of AI computation and roughly $100,000 in API costs. (Anthropic)
- [● 5 SOURCES] Neither discovered attack currently threatens production cryptographic systems. (Anthropic)