Anthropic reported that its Claude Mythos Preview model helped improve attacks on HAWK, a post-quantum signature candidate, and on reduced-round AES. The findings are notable for AI-assisted cryptanalysis, but the cited sources do not show a practical break of Bitcoin or of production AES.
Anthropic reported that Claude Mythos Preview helped discover new cryptanalytic shortcuts against experimental and reduced-round cryptographic targets.
In a research post titled “Discovering cryptographic weaknesses with Claude,” Anthropic said its Claude Mythos Preview model contributed to improved attacks on HAWK, a lattice-based digital signature scheme, and on reduced-round versions of AES. Anthropic framed the work as evidence that frontier language models can assist expert cryptographers in finding non-obvious weaknesses.
The headline HAWK result concerns HAWK-512, a post-quantum signature candidate. According to Anthropic, Claude Mythos helped identify a “nontrivial automorphism” that weakens the scheme’s effective security and makes secret-key recovery substantially easier than previously known. CyberScoop, reporting on the work, described the result as weakening HAWK’s effective key strength while emphasizing that HAWK is not widely deployed in real-world systems.
Anthropic also said Claude Mythos improved a theoretical attack on seven-round AES. That distinction matters: standard AES uses more rounds than the reduced-round target studied in the research. Anthropic’s own summary says the AES finding does not break full AES and does not require production software changes.
HAWK is part of the broader post-quantum cryptography landscape: schemes designed to remain secure even if large, fault-tolerant quantum computers eventually undermine classical public-key systems. A weakness in such a scheme is important because post-quantum systems are being evaluated as future replacements for cryptographic primitives that are vulnerable to quantum attacks.
The Anthropic result is therefore a warning about the difficulty of designing and vetting post-quantum cryptography. The company said HAWK had already received expert attention before Claude Mythos found a new shortcut. CyberScoop similarly characterized the finding as a meaningful cryptanalytic result rather than a mere benchmark exercise.
Anthropic has also published a public GitHub repository, “anthropics/cryptography-research-demo,” containing research artifacts and cryptanalysis code folders related to HAWK, AES, and LEA. A related arXiv paper, “CryptanalysisBench: Can LLMs do Cryptanalysis?,” coauthored by Anthropic researchers and academic partners, presents cryptanalysis as a high-stakes evaluation area for language models and reports that frontier models can produce novel cryptanalytic results under research conditions.
The available sources do not show that Claude broke Bitcoin’s signature scheme or produced a practical attack on Bitcoin wallets. Bitcoin currently uses elliptic-curve digital signatures, which are not designed to be secure against a sufficiently capable quantum computer. That is a separate, long-discussed risk: a powerful enough quantum computer could derive private keys from exposed public keys under current elliptic-curve assumptions.
The Anthropic HAWK result is relevant to Bitcoin only indirectly. If Bitcoin or other systems eventually migrate to post-quantum signatures, the security of candidate schemes will be critical. Anthropic’s work suggests AI systems may become useful tools for finding weaknesses during that evaluation process. It does not establish that all post-quantum signatures are unsafe, nor that Bitcoin faces an immediate AI-enabled cryptographic break.
Likewise, the AES result should not be read as a break of modern encrypted systems. Anthropic and CyberScoop both describe it as a theoretical improvement against reduced-round AES, not full AES. Anthropic explicitly says the AES result does not require changes to production software.
The most grounded conclusion is narrower than some social media interpretations: Anthropic has demonstrated that an advanced language model can assist in discovering meaningful cryptographic weaknesses in selected research targets. That is significant for cryptographers and security teams because it may accelerate both offensive analysis and defensive review.
But the cited evidence does not support claims that Bitcoin is newly compromised, that full AES is broken, or that deployed post-quantum cryptography has broadly failed. The findings are best understood as a serious research milestone in AI-assisted cryptanalysis—and as another reason for careful, public, expert review before cryptographic schemes are deployed at scale.
Anthropic reported that Claude Mythos Preview helped discover new cryptanalytic shortcuts against experimental and reduced round cryptographic targets.
Anthropic framed the work as evidence that frontier language models can assist expert cryptographers in finding non obvious weaknesses.
The headline HAWK result concerns HAWK 512, a post quantum signature candidate.
Continue reading
