Agentic cryptanalysis: the threshold where Claude Mythos Preview rewrites the crypto-review stack

On 28 July 2026, Anthropic's Frontier Red Team published a result that moves the frontier: Claude Mythos Preview is no longer limited to spotting implementation bugs in crypto libraries. Per the research post, the model found mathematical flaws in the algorithms themselves — a category shift for anyone wiring multi-agent security harnesses, PQC migration plans, or red-team pipelines.

Where the field actually sits today

Anthropic reports two concrete findings. First: an improved attack on HAWK, a post-quantum digital signature still in NIST's third-round Additional Digital Signatures track. In a multi-agent harness with Python and Sage, Mythos Preview surfaced a previously unexploited nontrivial automorphism in HAWK's lattice. Effective keysize is cut roughly in half. For HAWK-256, Anthropic says the expected full key-recovery cost moved from 2⁶⁴ to 2³⁸. The attack stays exponential and is specific to HAWK — it does not hit other NIST PQC signature candidates or lattice crypto broadly. HAWK is not deployed in production.

Second: a stronger meet-in-the-middle attack on a research-only AES-128 reduced to 7 of 10 rounds. Mythos designed a fingerprint it named the Möbius Bridge, removing a 256-value guess stage. Depending on runtime measurement, Anthropic reports a 200–800× speedup — under an academic chosen-plaintext model (~2¹⁰⁵ plaintexts). Full-round production AES is unaffected.

Each track cost roughly $100,000 in API usage. HAWK took about 60 hours, with a human operator who was not a lattice specialist and mostly did project management. AES discovery ran largely autonomously via a scaffold. Anthropic followed responsible disclosure (HAWK authors, NIST list, government and industry partners), released demo code plus two technical papers, and co-built CryptanalysisBench with researchers at ETH Zurich, Tel Aviv University, and TU Berlin so others can score LLM cryptanalysis capability.

Three trajectories that look highly likely in 6–12 months

1. Post-quantum standard review will treat LLM agents as co-reviewers. Anthropic states that reviewing specs like HAWK with AI is a powerful pre-deployment tool. Highly likely: agentic campaigns run in parallel with human expert review on future standardization rounds.
2. The bottleneck shifts from discovery to human verification. For AES, the model produced the attack in about a week; researchers then spent hundreds of hours validating mathematical correctness. In 6–12 months, labs without a proof-checking pipeline will saturate first.
3. Crypto red-team stacks standardize on multi-agent harnesses plus computer algebra. The published setup — collaborating workers, sandbox, Python/Sage, literature access — is already a blueprint. Plausible that internal security teams replicate it against their own primitives and PQC migration candidates.

What stays speculative: attacks that break live production systems. Anthropic is explicit: no production software has to change because of these two results.

Three capabilities to lock in this quarter

• A reproducible agentic cryptanalysis harness — multi-worker, sandboxed, formal-math tooling, and logs of rejected hypotheses (Anthropic notes one worker dismissed the winning idea early; a second pushed it through).
• A model-independent math validation chain — human review, formalization, regression checks — before any “critical finding” ticket leaves the lab.
• An inventory of PQC candidates vs research-only reduced ciphers — so academic signal never gets mistaken for a production incident.

Three risks to mitigate now

• Headline overreach: “AI breaks AES” is not what Anthropic published. The result targets 7-round AES, not the 10-round production cipher.
• Publish-first, prove-later culture: if discovery cost (~$100k API, tens of hours) falls faster than proof cost, false-positive noise will flood security teams.
• Agent harness side channels: a setup wired to literature, code, and poorly isolated test keys can leak more than a paper result. Isolation and responsible disclosure remain non-negotiable — Anthropic shared the HAWK attack with the scheme's authors in June before public release.

Three levers to activate this week

1. Read Anthropic's post plus the HAWK and Möbius Bridge papers; clone anthropics/cryptography-research-demo and re-run the local verification path.
2. Wire CryptanalysisBench (arXiv:2607.18538) into your internal eval grid — not for leaderboard vanity, to calibrate what your agent stack can actually prove.
3. Update the post-quantum migration runbook: any NIST-track candidate gets an “agent cryptanalyst + human validation” gate before product integration.

Is your crypto-review pipeline still 100% human?

If you're into the latest AI-driven tech, I publish a deep dive every day on frontier models, hardware, robotics, automations and AI-generated music. 👉 Get the next one straight in your inbox — sign-up takes ten seconds.

Sources

• Discovering cryptographic weaknesses with Claude (Anthropic)