Claude AI discovers vulnerabilities in cryptographic algorithms
Anthropic reported that its Claude Mythos Preview AI model discovered mathematical weaknesses in the HAWK post-quantum signature algorithm and improved attacks on simplified AES-128 encryption. Anthropic said both findings do not affect currently operational systems and coordinated disclosure with the US government and the HAWK developers.
What happened
Claude Mythos Preview discovered mathematical weaknesses in HAWK, a post-quantum digital signature algorithm candidate in NIST's standardization plan.
Using a multi-agent semi-autonomous system, Claude identified previously undetected symmetry in the mathematical structure of HAWK, enabling faster key recovery attacks. In approximately 60 hours and costing about 10 million USD in API usage, Claude reduced the computational complexity for the smallest HAWK-256 variant from 2^64 operations to 2^38 operations, demonstrating the key size cannot provide its claimed security strength.
Claude also developed a new attack method on a simplified 7-round AES-128 encryption algorithm, improving attack speed by 200 to 800 times over prior best methods. The complete AES-128 has 10 rounds, so this result does not break currently deployed AES.
Reported impact
Anthropic stated that both findings do not currently affect operational systems, and Claude Mythos Preview remains unavailable to the public.
The HAWK vulnerability finding may require key size doubling to restore claimed security levels, potentially compromising the algorithm's design advantage of smaller keys and affecting its candidacy for NIST standard selection.
Index analysis
Claude Mythos deployed a multi-agent approach where different AI instances reviewed cryptographic literature, proposed attack methods, and performed computational verification, with human researchers managing workflow rather than providing direct solutions.
This marks the first time Claude identified new attack methods directly from algorithmic mathematical design rather than implementation errors in cryptographic software.
Governance controls
Anthropic coordinated vulnerability disclosure with HAWK developers before public announcement, informing the HAWK team in June 2026.
Anthropic disclosed findings through NIST public email forums and collaborated with institutions including ETH Zurich on developing the CryptanalysisBench benchmark.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.