TL;DR
- Attack Improvements: Claude Mythos Preview has improved attacks on HAWK and seven-round AES-128 without breaking deployed cryptography.
- Measured Gains: The HAWK recovery cost fell from 2^64 to 2^38 operations, while the AES attack became 200 to 800 times faster.
- Production Limits: HAWK is not deployed, and production AES-128 uses ten rounds rather than the seven tested.
- Human Control: Human operators reportedly sustained the search, while a 191-task benchmark and public code enable further testing.
Anthropic researchers used Claude Mythos Preview to improve attacks on HAWK and seven-round AES-128 on July 28. HAWK is a candidate digital-signature scheme intended for quantum-era threats, while the Advanced Encryption Standard (AES) experiment used a deliberately reduced research variant. Neither result compromises deployed systems: HAWK is not deployed, and production AES-128 uses ten rounds.
Anthropic says that about 60 hours of Claude-assisted work improved the HAWK attack after the candidate had undergone two years and two rounds of expert review. The company estimates that each of the two main findings consumed roughly $100,000 in API usage. Faster pre-deployment evaluation could help standards reviewers reject weak parameters earlier, although the cost and continued expert control keep the method far from a push-button replacement for cryptographers.
Cryptanalysis searches for mathematical shortcuts that reduce the work needed to defeat a cryptographic design. A researcher prompting Claude Mythos repeatedly directed it to pursue difficult, publishable attacks instead of switching targets. Because human direction supplied the objective and persistence, the experiment measures AI-assisted research rather than a fully autonomous cryptographer.
What Claude Found
Claude Mythos reduced the demonstrated expected cost of HAWK-256 full-key recovery from 2^64 to 2^38 operations. It identified a previously unused symmetry in HAWK’s underlying lattice structure. Exploiting that symmetry narrowed the candidates the attack had to enumerate.
Halving HAWK’s effective key strength means HAWK key sizes would need to double to keep the same level of security. Standards reviewers have a concrete parameter change to examine before approval or deployment.
For AES-128, Claude eliminated one attacker guess from a meet-in-the-middle attack, making earlier attacks 200 to 800 times faster. Meet-in-the-middle cryptanalysis computes forward from one side of an encryption process and backward from the other until intermediate values match. Claude’s experiment covered seven rounds, not the ten transformations used by production AES-128.
Claude’s attack also assumes 2^105 chosen plaintexts, meaning messages selected by an attacker for analysis, or more than 400 octillion messages. Generating that input leaves no practical route to production keys or software.
HAWK’s lower recovery cost forces reviewers to reconsider a candidate parameter before standardization. Claude’s AES acceleration improves a constrained research attack but stops three rounds and an impractical volume of attacker-selected input short of the deployed cipher. Independent cryptographers have not yet reproduced either result, so the work does not establish broad autonomous capability.
Human Steering Sets the Boundary
Anthropic researchers repeatedly told the model to persist, reject easier targets, and seek work suitable for publication. During the 60-hour HAWK run, one Claude worker rejected the lattice-symmetry idea as infeasible; a second exploited it, shared the method across the agent group, and helped reduce expected recovery to 2^38 operations. Human operators still set the standard of success, rejected easier alternatives, and decided when the multi-agent system should continue searching.
Anthropic shared the HAWK attack with its authors in June and coordinated public disclosure through the US National Institute of Standards and Technology (NIST) mailing list. HAWK’s designers can evaluate the weakness before a standardization or deployment decision rather than after affected software ships.
Claude Mythos Preview became available for advanced computer-security work in March. Its earlier software-vulnerability research targeted flaws in code; HAWK and AES demand mathematical shortcuts through cryptographic designs. High API expense, consequential prompting, and restrictive assumptions keep the AES method outside production conditions.
CryptanalysisBench contains 191 tasks across six cryptographic families, largely drawn from four NIST standardization competitions. Its breadth can test whether performance extends beyond two carefully pursued attacks, but benchmark results alone cannot independently verify the new HAWK and AES methods.
Anthropic published a public cryptanalysis code repository accompanying the associated papers. Anthropic’s demonstration material allows specialists to inspect the HAWK lattice technique and reduced-round AES method. Standards reviewers can use the repository to test the 2^38 HAWK-256 recovery before deciding whether HAWK needs larger keys.