Anthropic’s Mythos Preview Finds Cryptographic Weaknesses in HAWK and Reduced-Round AES
On July 28, 2026, Anthropic’s Frontier Red Team reported that Claude Mythos Preview discovered an improved key-recovery attack on the HAWK post-quantum signature candidate and a 200–800× faster attack on seven-round AES-128—mostly autonomously, at roughly $100,000 API cost each—with no impact on production systems today.
TLDR
Anthropic on July 28, 2026 published “Discovering cryptographic weaknesses with Claude,” reporting that Claude Mythos Preview found improved mathematical attacks on two targets: HAWK (a NIST post-quantum signature candidate) and a seven-round reduced variant of AES-128. Results were developed mostly autonomously (~$100,000 API cost each; HAWK ~60 hours), disclosed to algorithm authors, NIST channels, and U.S. government/industry partners. Anthropic stresses neither result breaks production systems today.
The two primary results
| Target | What Mythos improved | Production impact |
|---|---|---|
| HAWK-256 (NIST PQC additional-signature candidate) | End-to-end key-recovery via a previously unexploited lattice automorphism; effective key strength roughly halved (e.g. small parameter cost cited as 2^64 → 2^38) | None deployed — HAWK is a candidate, not a fielded standard; larger parameters remain impractical to attack under this exponential improvement |
| 7-round AES-128 (research-only reduced cipher; full AES-128 has 10 rounds) | New meet-in-the-middle fingerprint (“Möbius Bridge”) claimed 200–800× faster than prior best under chosen-plaintext assumptions | Does not break full AES; prior-style attacks already assume huge chosen-plaintext budgets |
Anthropic also notes preliminary work on reduced-round LEA, Serpent, and limited gains on Salsa20, Poseidon, and SHA-1, with fuller write-ups planned after verification.
How the research was run
- HAWK: semi-autonomous multi-agent harness (Claude Code–like) with occasional human project management; human operator not a lattice specialist; end-to-end verification pipeline included.
- AES: scaffold for hypothesis → experiment → attack design; model initially refused (“AES is too studied”); a handful of nontechnical prompts over days, then roughly one billion output tokens refining the attack; humans spent hundreds of hours validating.
- Artifacts: technical PDFs (HAWK + AES), AES chain-of-thought extract, demo code, and CryptanalysisBench (with ETH Zurich, Tel Aviv University, University of Haifa / TU Berlin academic partners per Anthropic posts).
Why this story matters
This is a capability milestone for AI-assisted cryptanalysis, not a “your bank is broken” headline. The durable points: (1) frontier models can contribute publishable crypto research under human verification bottlenecks; (2) NIST-style public contests are exactly where such attacks belong before deployment; (3) verification cost is becoming the scarce resource. Expect more AI-found weaknesses in long-tail ciphers—and pressure on labs and governments for disclosure norms when attacks do hit live systems.