Anthropic’s LLM Claude Mythos contributed to cryptanalytic research by identifying a key-recovery attack on HAWK-512, a post-quantum signature scheme candidate, reducing its estimated security from 128 bits to at most 108 bits and possibly as low as 81 bits, according to Appendix C of the paper. Although this attack is not practical for the full scheme, it represents a meaningful theoretical improvement over previous analyses. The same LLM also produced an improved key-recovery attack on 7-round AES-128, which, while not threatening the full 10-round AES-128, is notable given the rarity of new cryptanalytic results for AES. Importantly, Mythos found no stronger attack against full-round AES, reinforcing confidence in its resilience. Anthropic also helped develop CryptanalysisBench, a benchmark for evaluating LLMs on cryptanalysis tasks across AES, ChaCha, BLAKE3, and post-quantum schemes. Despite these advances, the author expresses confidence that LLMs will not break established symmetric cryptographic schemes. This confidence stems from several factors: symmetric ciphers like AES, SHA-3, and BLAKE3 are designed with high-level structural security through modes such as GCM, CTR, and sponge constructions; they intentionally lack mathematical structure to resist simplification and exploitation; their security largely depends on differential cryptanalysis, a well-explored field where entirely new attack classes are unlikely to emerge; and these algorithms have undergone extensive, battle-tested scrutiny over thousands of hours, mostly unpublished, showing increased security with each additional round. The author notes that LLM-assisted research is better directed toward finding bugs in security proofs or underanalyzed areas like post-quantum standardization candidates, rather than attempting to break robust symmetric primitives.
Key facts
- Anthropic’s LLM Claude Mythos discovered a key-recovery attack on HAWK-512 reducing its security from 128 bits to at most 108 bits, possibly as low as 81 bits
- The same LLM produced an improved key-recovery attack on 7-round AES-128 but found no stronger attack against full-round AES-128
- Anthropic helped create CryptanalysisBench, an LLM benchmark for cryptanalysis tasks on AES, ChaCha, BLAKE3, and post-quantum schemes
- Symmetric ciphers derive security from high-level construction modes (e.g., GCM, CTR, sponge) and intentional lack of mathematical structure
- Symmetric cryptanalysis largely relies on differential cryptanalysis, a field extensively explored by humans over thousands of hours