Anthropic researchers have revealed that its Claude Mythos Preview model autonomously identified significant mathematical flaws in two major cryptographic systems, advances that expert human review failed to catch despite years of scrutiny.
The first breakthrough targeted HAWK, a post-quantum digital signature scheme currently under third-round consideration in NIST’s standardization process.
Mythos discovered a previously unidentified “nontrivial automorphism” in HAWK’s underlying lattice structure, effectively halving the scheme’s key strength.
Claude AI Discovers Cryptographic Weaknesses
This cut the expected attack cost for HAWK-256 from operations down to , a result that emerged after HAWK had already survived two rounds of expert human review over two years.
The second finding involved a novel technique the model called the Möbius Bridge, applied against a 7-round reduced version of AES-128, the world’s most widely used symmetric cipher.
This fingerprinting method eliminated one exhaustive guessing stage from the best-known meet-in-the-middle attacks, improving their speed by 200 to 800 times.
The attack remains entirely theoretical, requiring chosen plaintexts, and does not touch the full 10-round cipher used in real-world systems.
Anthropic was careful to note that neither result has any practical impact on current production software. HAWK has not been deployed anywhere, and the AES attack only works against a deliberately weakened variant of the cipher.
Both results emerged from largely autonomous AI workflows with minimal human steering. The HAWK attack took about 60 hours of semi-autonomous work, with a human researcher providing project management rather than technical cryptography expertise.
The key insight came from two AI worker agents collaborating on the same idea, where one agent initially dismissed the approach as infeasible before the other found a way to fully exploit it.
The AES discovery followed a more unusual path. Mythos initially refused to engage with the problem, insisting that AES was “genuinely hard” and that there was “nothing easy to find” in the most heavily studied cipher in existence.
Researchers responded with informal, typo-laden prompts pushing the model to search for genuinely novel ideas rather than settling for easy results.
In response, Claude rewrote its own experimental harness and, over the following three days and roughly one billion output tokens, independently arrived at the Möbius Bridge technique. Each of the two main results cost approximately 100,000 dollars in API compute to produce.
Anthropic also disclosed several smaller preliminary results, including a practical attack on 13-round LEA, a lightweight cipher used in ISO-standardized systems, along with modest improvements against Serpent-128, Salsa20, Poseidon, and SHA-1.
None of these pose any immediate real-world risk, but they point to AI models steadily expanding their reach into serious cryptanalysis.
To help the wider research community track this trend, Anthropic partnered with ETH Zurich, Tel Aviv University, and the University of Haifa to release CryptanalysisBench, a benchmark designed to evaluate how well large language models perform against a broad set of cryptographic ciphers.
Anthropic followed responsible disclosure procedures throughout, coordinating with HAWK’s authors and briefing government and industry partners before going public.
The company frames these findings as adversarial cryptographic review working exactly as intended, while cautioning that as AI cryptanalysis capabilities grow, the field will need to prepare for scenarios where a model uncovers a flaw in a cryptosystem that actually matters in production, a challenge that may soon outpace the human capacity to verify and respond.
Cut SOC investigation blind spots and contain threats earlier to reduce response costs and business disruption with ANY.RUN.
