AI produces its first original cryptanalysis results @AnthropicAI's Frontier Red Team published two cryptanalysis results on Tuesday from Claude Mythos Preview, marking a qualitative shift in
@AnthropicAI's Frontier Red Team published two cryptanalysis results on Tuesday from Claude Mythos Preview, marking a qualitative shift in what AI can do in advanced mathematical research. Unlike the implementation bugs the model surfaced in April, these are attacks on the algorithms themselves.
The first result targets HAWK, a lattice-based digital signature scheme currently under evaluation by the US National Institute of Standards and Technology (NIST) as a post-quantum cryptography standard. According to Anthropic, despite HAWK having survived two rounds of expert human review over roughly two years, Mythos was able to improve the best-known attack on it in just 60 hours, effectively cutting its key strength in half.By Anthropic's own figures, cracking the smallest HAWK key fell from 2^64 steps to 2^38, roughly 67 million times less work.
HAWK is one of nine algorithms that NIST advanced to the third round of its post-quantum cryptography standardization process, as part of a broader effort to prepare cybersecurity systems for future quantum computing threats.HAWK advanced to the third round, although the agency had already encouraged additional analysis of its underlying security assumptions.
The second result involves AES, one of the most widely used encryption standards in the world. Researchers used Claude to improve attacks against a reduced version of AES. Specifically, Claude accelerated a known attack on a seven-round version of AES-128 by 200 to 800 times. Full AES runs ten rounds and remains unaffected.
Anthropic says it believes the attack discovered by Mythos Preview does not impact the other NIST post-quantum cryptographic schemes or other schemes that use related methods.Anthropic emphasized that neither discovery creates an immediate risk for production systems or currently deployed encryption. Nothing in either result touches SHA-256 or ECDSA, the cryptographic primitives underpinning Bitcoin and most existing blockchain infrastructure.
Anthropic says Claude developed the HAWK attack in roughly 60 hours and estimates each breakthrough cost about $100,000 in API compute. A year ago, AI models could not perform original cryptanalysis of this kind at all. The findings show how artificial intelligence is moving beyond traditional software development tasks and entering advanced mathematical research.
Anthropic: Discovering cryptographic weaknesses with Claude | The Quantum Insider: NIST Advances Nine Post-Quantum Signature Algorithms to Third Round | BeInCrypto: Anthropic's AI Finds Weaknesses in the Technology Protecting Your Data
