Claude AI Just Cracked a Post-Quantum Test Scheme and Found a Faster 7-Round AES Attack — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic said Claude Mythos Preview helped derive an end-to-end key-recovery attack against HAWK-256, with the released implementation running in roughly 3 hours 42 minutes on a 96-core server and dropping the expected work factor from 2^64 to 2^38.
- Anthropic said Mythos found an invariant it calls the Möbius Bridge that lets the attack skip a 256-way enumeration step, yielding a 200- to 800-fold speedup on AES-128 reduced to seven rounds — though the attack still demands about 2^105 chosen plaintexts.
- HAWK remains a third-round candidate in NIST's post-quantum signature process as of late July 2026, and Anthropic's released code supports only the HAWK-256 challenge parameter, not the HAWK-512 or HAWK-1024 security-level sets, which Anthropic says remain impractical to attack despite lower gate-count estimates.
- Anthropic said Mythos largely conducted both projects itself: the HAWK analysis ran about 60 hours in a multi-agent setup at roughly $100,000 in API cost, while the AES work generated about one billion output tokens over several days.
- Anthropic disclosed the model initially refused to work on AES, insisting improving on the cipher was impossible, and published the blunt follow-up prompts that pushed it past the refusal — with human verification taking nearly a month and emerging as the visible bottleneck.
- The disclosures follow the July 20 release of CryptanalysisBench, a 191-task benchmark developed with ETH Zurich, the University of Haifa, TU Berlin and Tel Aviv University, where models including Mythos 5 broke 65–86% of easier first-tier schemes and 6–12 full-strength second-tier ones.
Why it matters: Two AI-led cryptanalysis wins that stop short of breaking anything deployed: HAWK-512 and HAWK-1024 remain impractical, and the seven-round AES attack needs an unattainable plaintext budget, so NIST's third-round slate and fielded encryption are unaffected. The workflow shift is the story — a roughly $100,000 model run plus months of human checking produced peer-grade papers, positioning Mythos as a research labor multiplier for cryptanalysts rather than a replacement.
Ask SkimNews
