Anthropic Demonstrates LLM-Assisted Cryptanalysis with Claude Mythos, Finds New Attacks on HAWK and AES
Key Takeaways
- ▸Anthropic used Claude Mythos to discover new cryptanalytic attacks, including a key-recovery attack on HAWK post-quantum scheme and improved attacks on reduced-round AES
- ▸The attacks demonstrate LLM capability in specialized security domains but pose no practical threat to deployed cryptography
- ▸Anthropic co-created CryptanalysisBench, an LLM benchmark for cryptanalysis, to enable standardized evaluation of AI capabilities in this field
Summary
On July 28th, Anthropic announced new cryptanalytic attacks discovered using their LLM Claude Mythos, demonstrating the emerging capability of large language models in specialized security research. The breakthroughs include a key-recovery attack on HAWK, a post-quantum signature scheme candidate in NIST's standardization project, which reduces HAWK-512's estimated security from 128 bits to at most 108 bits, and an improved key-recovery attack on 7-round AES-128. Notably, Mythos found no stronger attacks against full-round AES, and the attacks themselves are not practically exploitable against production cryptographic systems.
Beyond the specific attacks, Anthropic helped develop CryptanalysisBench, a new LLM benchmark comprising cryptanalysis tasks on both full and reduced-round versions of AES, ChaCha, BLAKE, and other algorithms submitted to NIST's competitions. The company's work includes a call for the cryptography community to formalize attack techniques and develop more unified descriptive frameworks across security research. The research underscores an important insight: while LLMs can accelerate cryptanalytic research and uncover novel attacks, the inherent messiness and robustness of symmetric cipher design makes it highly unlikely that language models will break established cryptographic schemes in use today.
- LLMs may accelerate cryptographic research and uncover errors in security proofs, but the deliberate messiness of cipher design makes breaking established algorithms fundamentally unlikely
Editorial Opinion
This research is significant not because it threatens AES or SHA-3—it doesn't—but because it demonstrates how LLMs can be productively applied to specialized technical domains and accelerate domain expertise. The broader contribution may be Anthropic's call for formalizing cryptanalytic techniques; too much of the field relies on ad hoc notation and reinvented frameworks when researchers are fundamentally solving the same problems. Whether LLMs break symmetric crypto matters less than whether they help the field become more rigorous.
