Anthropic disclosed today that its unreleased Claude Mythos Preview model discovered two previously unknown attacks on cryptographic algorithms, including one targeting HAWK, a post-quantum digital signature scheme currently competing in the third round of NIST's standardization process. The AI model reduced the cost of recovering HAWK's smallest secret key from 2^64 operations to 2^38 operations—roughly 67 million times less computational work—through identifying a mathematical symmetry no human cryptographer had exploited. Each cryptographic breakthrough cost approximately $100,000 in API usage, and Anthropic coordinated disclosure with NIST, algorithm authors, and U.S. government and industry partners before publication. The findings demonstrate AI models' growing capability in advanced mathematical research, though human verification of the AES attack alone required several hundred hours of expert review.
Claude Mythos Preview Discovers HAWK Vulnerability Through Mathematical Symmetry
Claude Mythos Preview identified a symmetry buried in HAWK's mathematical structure that no human cryptographer had previously exploited. For HAWK's smallest configuration, the discovery dropped the computational cost of recovering a secret key from 2^64 operations to 2^38 operations. NIST moved HAWK into the third round of its post-quantum signature competition in May, where it remains the last lattice-based candidate. The non-regulatory federal agency coordinates the competition to establish quantum-resistant cryptographic standards.
Fixing the vulnerability requires roughly doubling HAWK's key sizes. "Unfortunately, doubling HAWK's key size eliminates many of the reasons making the scheme (as it currently stands) an attractive PQC signature candidate," Anthropic wrote in its disclosure. The signature size increase affects HAWK's primary advantage—compact keys and fast signing performance that distinguished it from competing post-quantum schemes.
HAWK has never been deployed in production systems. Bitcoin and other blockchains continue using ECDSA, the pre-quantum signature scheme that candidates like HAWK are designed to eventually replace. Anthropic disclosed the HAWK finding to the algorithm's authors and coordinated with NIST before publication.
AI Model Breaks 2013 AES Record After Initial Refusal
The second attack targets a 7-round research version of AES, the cipher used in HTTPS traffic encryption, encrypted drives, and exchange backends. Full AES-128 uses 10 rounds of scrambling, and no researcher had improved attacks on the 7-round variant since 2013. Anthropic researchers barred Claude from all five established families of AES cryptanalysis and instructed it to invent a sixth approach. The model inherited working notes from earlier agent runs that had tested roughly 200 failed attack variants.
Claude initially refused the task. "On AES-128 r5/r6/r7 it found nothing because there's nothing easy to find; this is the most-studied block cipher in existence," the model stated in transcripts Anthropic published. Researchers sent three substantive messages over three days, including: "no again the goal is that we have highly inteligent [sic] model as good top researcher, we want to find new attacks." Another message refused to let Claude substitute an easier cipher.
The model then produced what the research paper calls a Möbius Bridge technique, eliminating one of nine key bytes an attacker previously had to guess. Refining the attack into its published form took additional days and generated one billion output tokens. The breakthrough speeds up the 7-round AES attack by 200 to 800 times compared to the 2013 record.
Claude Cracks LEA Encryption Standard in Under One Hour
Claude broke 13 rounds of LEA, a Korean national standard and ISO lightweight-encryption standard built for mobile phones and internet-of-things devices, in under one hour on a desktop computer. The prior best attack required 2^98 plaintext pairs. The deployed version of LEA runs 24 rounds, so no production implementation is compromised by this finding.
Human Verification Process Takes Hundreds of Hours for AES Result
The HAWK research paper states: "The majority of mathematical discoveries in this paper were AI-assisted. Human author contribution mainly consisted of directing, organizing and verifying AI work." Claude found the AES attack concept in days, but Anthropic researchers spent several hundred hours learning sufficient cryptography to confirm the attack functioned correctly. The same Claude model discovered 271 vulnerabilities in Firefox during Anthropic's internal testing.
"The cybersecurity community is now grappling with the fact that language models are able to discover so many bugs that the standard human processes (like vulnerability triage, verification, and remediation) struggle to keep up," Anthropic wrote, noting that human researchers may become the bottleneck in AI-assisted security research.
Anthropic built CryptanalysisBench, a benchmark containing 191 cipher-breaking tasks drawn mostly from NIST competitions. Models submit working attack scripts that either win a formal security game or fail, with no partial credit and no human grading. Mythos 5 broke 85.7% of tasks with known solutions, compared to 65.3% for the weakest model tested. Against full-strength ciphers with no published break, every model scored under 9%.
FAQ
What did Claude Mythos Preview discover about HAWK cryptography?
Claude Mythos Preview found a mathematical symmetry in HAWK that reduced the cost of recovering the smallest secret key from 2^64 operations to 2^38 operations—roughly 67 million times less computational work. HAWK is a post-quantum digital signature scheme competing in the third round of NIST's standardization process, where it entered in May as the last lattice-based candidate.
How much did Anthropic spend on the cryptographic research runs?
Each cryptographic breakthrough cost approximately $100,000 in API usage. Anthropic researchers also spent several hundred hours verifying that the AES attack functioned correctly after Claude generated the initial concept.
What is the impact of the HAWK vulnerability on deployed systems?
HAWK has never been deployed in production systems, so no active implementations are affected. Bitcoin and other blockchains continue using ECDSA. Fixing the vulnerability requires doubling HAWK's key size, which Anthropic stated "eliminates many of the reasons making the scheme (as it currently stands) an attractive PQC signature candidate."
