Anthropic's Claude Mythos Preview has uncovered cryptographic vulnerabilities in HAWK and AES-128, although they pose no immediate threat to existing systems.
Anthropic has revealed that its Claude Mythos Preview has played a pivotal role in identifying two significant cryptanalytic attacks: one on the post-quantum digital signature scheme HAWK and another on a reduced version of AES-128. The company shared this information recently.
"Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms—the mathematical methods that are used to keep data private." — Anthropic
While the findings are noteworthy, Anthropic clarified that they do not pose immediate risks to existing systems, emphasizing that no changes to production software are required at this time.
HAWK is a candidate for standardization as a post-quantum digital signature, crucial for verifying message or transaction authorship without exposing private keys. Conversely, AES-128 is a widely utilized symmetric cipher that employs a single secret key for both encryption and decryption. The findings relate specifically to a 7-round version of AES-128, which is used in academic assessments of resilience.
Each study incurred an estimated cost of around $100,000 in API usage, with the research and validation for HAWK taking approximately 60 hours, while the AES findings required several hundred hours of effort.
HAWK Vulnerability and Key Recovery
In its analysis, Anthropic noted that Claude Mythos Preview identified new symmetry within HAWK's mathematical structure, facilitating a more efficient key recovery attack. The effective resilience of keys in HAWK-256 was reduced, decreasing the expected key recovery cost from 2^64 to 2^38 operations. However, for larger key sizes in HAWK-512 and HAWK-1024, the attack remains practically unfeasible.
To preserve the original security level of HAWK, key sizes would need to be significantly increased, impacting the scheme's compactness. Anthropic underlined that these vulnerabilities do not extend to other NIST candidates or indicate a broader compromise in lattice-based post-quantum cryptography.
AES-128 Research Findings
Anthropic's second finding pertains to the 7-round AES-128, leveraging a method called the Möbius Bridge. This approach markedly accelerated the previously known meet-in-the-middle attack, increasing efficiency by 200 to 800 times, although it remains impractical under real-world conditions.
Despite the significant acceleration, the attack requires 2^105 queries and does not compromise full AES-128. The research outcomes are more theoretical and do not threaten current operational systems.
Anthropic shared its findings with the developers of both HAWK and AES, as well as with U.S. government agencies and industry partners, ensuring transparency and collaboration within the cryptographic community.
New Benchmark for Cryptanalysis
In a move to evaluate the cryptanalytic capabilities of language models, Anthropic, in collaboration with researchers from several universities, has introduced CryptanalysisBench. This benchmark features 191 tasks across various cryptographic primitives derived from NIST competitions. The goal is for models to explain vulnerabilities and deliver working attack scripts that pass formal verification.
Initial results showed that advanced models solved between 65% and 86% of first-tier tasks, with Claude Mythos 5 achieving an 85.7% score.
Additional Findings and Future Work
Furthermore, Anthropic reported discovering an attack on the 13-round version of the Korean standard block cipher LEA, which can recover the key in under an hour on modern hardware. However, the full LEA employs 24 rounds, meaning this finding does not currently threaten operational systems.
Anthropic's ongoing research continues to yield insights into cryptographic vulnerabilities, with previous results indicating over 10,000 high and critical severity security issues identified through its vulnerability discovery program.
