On July 28, Anthropic published two cryptanalysis results produced with Claude Mythos Preview. One concerns HAWK, a post-quantum digital-signature candidate in the third round of a National Institute of Standards and Technology evaluation. The other improves a research attack on a reduced-round version of AES, the symmetric cipher used across modern computing.
The headline needs careful boundaries. HAWK is a candidate under review, with no deployed standard at risk. The AES work applies to seven-round AES-128 under an impractical chosen-plaintext assumption; production AES-128 uses ten rounds. Anthropic says the results do not affect production AES, other NIST post-quantum candidates, or lattice cryptography as a category.
Within those boundaries sits a consequential signal. AI can now contribute novel attack ideas to the adversarial review that makes cryptographic standards trustworthy. Discovery moved quickly. Verification, disclosure, and institutional judgment still required expert labor.
A result inside the process
Anthropic reports that Mythos identified the central idea behind the HAWK attack in roughly 60 hours, with occasional project-management guidance from a human researcher. HAWK had already passed through two rounds of public expert review over about two years. The model found a previously unused symmetry in the mathematical structure behind HAWK, reducing the effective security of one parameter set. Anthropic says larger HAWK parameter sets remain impractical to attack.
This is how a candidate process earns credibility. NIST lists HAWK among nine Round 3 additional-signature candidates and provides public channels for technical comment. Anthropic says it notified HAWK's designers in June and coordinated disclosure through the NIST mailing list. Independent cryptographers can now inspect the paper, reproduce the demonstration, challenge the assumptions, and feed their conclusions into a live standards decision.
The AES result shows a similar capability in a safer research setting. Mythos developed a mathematical shortcut that Anthropic calls the Möbius Bridge, making a prior theoretical attack on seven-round AES-128 between 200 and 800 times faster. The attack still assumes about 2^105 chosen plaintexts, far beyond a practical setting, and it does not reach full ten-round AES-128.
The discovery speed can obscure the work that follows. Anthropic estimates that each main result consumed about $100,000 in API usage. The AES project generated roughly one billion model-output tokens, followed by several hundred hours of human validation. Researchers had to reconstruct the useful idea, prove it, test code, check scope, prepare a paper, and coordinate disclosure. The model expanded the search surface. People had to turn a promising trace into a defensible claim.
That division of labor is likely to spread. CryptanalysisBench, a new preprint from researchers at Anthropic, ETH Zurich, Tel Aviv University, and the University of Haifa, proposes a benchmark for measuring language-model performance on cryptanalysis. It is an early research instrument rather than an official certification system. Its existence suggests that model-assisted review may become a repeatable input to standards work.
Verification bottleneck
Verification is becoming the scarce institutional function.
- What moved faster: A model surfaced attack ideas against a post-quantum signature candidate and reduced-round AES on a much shorter timeline than conventional expert review.
- Who has to verify: Independent cryptographers, scheme designers, standards bodies, security agencies, and affected vendors must reproduce the result, define its scope, coordinate disclosure, and decide what changes.
- What preserves trust: Public papers, runnable demonstrations, explicit limits, a disclosure trail, and a recorded standards decision matter more than a dramatic capability claim.
- What to watch next: Independent responses to the HAWK paper, any NIST status change, and whether standards bodies formalize AI-assisted cryptanalysis as one review input among many.
Opportunities
Where value may appear: the handoff between machine discovery and expert judgment.
Builders could create reproducibility packets that bind a claim to source code, test conditions, model and tool versions, failed attempts, reviewer sign-off, and disclosure status. Standards teams may need triage systems that separate plausible weaknesses from artifacts before scarce cryptographer time is committed. Organizations also need practical cryptographic inventories: which schemes protect which systems, who owns each dependency, and how a change would move through testing and deployment.
None of this calls for an emergency migration because of these two findings. Production AES remains unaffected, and HAWK remains an undeployed candidate. The practical orientation is steadier: keep trust infrastructure visible, preserve the ability to change it, and insist that machine-speed discoveries arrive with evidence humans can inspect.
This is public-interest orientation and builder idea fodder, not cybersecurity, legal, procurement, financial, or investment advice.
The strongest institutions will gain more than faster discovery. They will build a review window that can keep up.
Sources
- Anthropic: Discovering cryptographic weaknesses with Claude
- NIST CSRC: Post-Quantum Cryptography, Round 3 Additional Signatures
- NIST CSRC: Post-Quantum Cryptography Standardization
- CryptanalysisBench: Can LLMs do Cryptanalysis?
- CyberScoop: What Anthropic found when it turned Mythos loose on encryption algorithms