LLMs won't break symmetric crypto

Anthropic announced that their LLM Claude discovered new cryptanalytic attacks, including a meaningful (though impractical) break of the post-quantum signature scheme HAWK and an improved attack on 7-round AES-128, though neither poses a threat to currently deployed cryptography. The research highlights LLMs' potential value in formalizing cryptanalytic techniques and finding errors in security proofs, but the author is confident that LLMs will not break established symmetric cryptographic schemes like AES, ChaCha, or SHA-3. Anthropic also created CryptanalysisBench, a benchmark for evaluating LLM performance on cryptanalysis tasks across various algorithms.