Claude Mythos AI Finds Flaws in Post-Quantum and AES Variants

Anthropic has announced that its LLM, Claude Mythos, successfully discovered several cryptanalytic attacks, including a key-recovery exploit against a post-quantum signature scheme. While the results demonstrate the growing capability of AI to assist in breaking encryption, security experts maintain that these breakthroughs do not threaten the core symmetric ciphers currently securing global data.

The AI's most significant find was a key-recovery attack on HAWK, a post-quantum signature scheme. According to reports from bfswa.blog, Claude Mythos reduced the security level of HAWK-512 from 128-bit to at most 108 bits, with some speculative estimates suggesting it could be as low as 81 bits. Additionally, the model identified an improved key-recovery attack on 7-round AES-128. Researchers emphasize that this specific attack poses no threat to the full 10-round version of AES used in production environments. To further these capabilities, Anthropic helped develop CryptanalysisBench, a specialized benchmark for LLMs that includes tasks focused on AES, ChaCha, BLAKE, and various post-quantum schemes.

The Nature of Symmetric Defense

Symmetric cryptography—which encompasses block ciphers, stream ciphers, and hash functions—is intentionally designed to be "messy." Unlike some asymmetric systems that rely on clean mathematical structures, symmetric primitives avoid such patterns to prevent exploitation. Most successful attacks in this domain rely on differential cryptanalysis, an empirical process of searching for statistical biases in input-output patterns.

Matthew Green describes the design philosophy of symmetric ciphers as akin to a farmer dragging a tractor into quicksand and burying it under cement. He notes that these structures are designed to be quick and easy to apply, but deliberately messy and difficult to untangle, which creates a natural barrier for the pattern-recognition strengths of LLMs.

Industry Implications

These findings suggest that LLMs can significantly accelerate the discovery of weaknesses in new or reduced-round algorithms. However, the consensus among specialists is that established primitives like full-round AES, ChaCha, and BLAKE3 remain robust. The primary utility of AI in this field is likely to shift toward identifying errors in security proofs and cryptanalytic complexity estimates rather than breaking proven schemes.

Future Outlook

As AI-driven cryptanalysis evolves, the focus of security risks is expected to shift toward underanalyzed post-quantum candidates and implementation bugs. While the HAWK results are a warning for emerging standards, the industry will continue to monitor whether LLMs can move beyond reduced-round attacks to challenge full-scale symmetric implementations. For now, the fundamental design of the world's most common ciphers appears to withstand the current generation of AI assistance.