Detailed Analysis
Anthropic's cryptography research team has demonstrated that Claude Mythos Preview, an experimental variant of its frontier model, can uncover novel weaknesses in established cryptographic algorithms—work that historically has required specialized human expertise cultivated over years. The two key results are an attack that significantly weakens HAWK, a digital signature scheme designed to be resistant to quantum computers, and a new attack technique against round-reduced AES, the symmetric cipher that underpins the vast majority of encrypted data worldwide, from HTTPS traffic to disk encryption. Anthropic is careful to note that neither discovery threatens deployed systems today: HAWK is still a research candidate rather than a widely adopted standard, and the AES attack applies only to a reduced-round variant of the cipher rather than the full, hardened version used in practice. Still, the fact that an AI model contributed to advancing state-of-the-art cryptanalysis on both a post-quantum scheme and the world's most ubiquitous cipher marks a notable milestone in AI-assisted mathematical research.
The significance of these findings lies less in immediate practical risk and more in what they signal about AI's growing capability in highly technical, adversarial domains. Cryptanalysis is a field where progress is typically slow, incremental, and dependent on deep specialized knowledge of algebraic structures, probability theory, and computational hardness assumptions. That a language model could identify novel attack vectors—effectively doing the kind of creative, exploratory reasoning that human cryptographers perform—suggests that AI systems are beginning to move from being tools that assist with known techniques to genuine collaborators capable of original research contributions. This matters because cryptography sits at the foundation of digital trust: the algorithms protecting financial transactions, government communications, and personal data all rely on assumptions about computational hardness that are validated through exactly this kind of adversarial stress-testing.
The timing is also relevant given the ongoing global transition to post-quantum cryptography. As organizations like NIST work to standardize new algorithms resistant to quantum attacks, schemes like HAWK are undergoing intense scrutiny precisely because the stakes of getting post-quantum standards wrong are enormous—a flawed algorithm adopted at scale could leave critical infrastructure vulnerable for years before the weakness is discovered and patched. Anthropic's research suggests that AI models can now serve as an additional, scalable layer of red-teaming for these emerging standards, potentially catching weaknesses earlier in the standardization process than would otherwise be possible. This has a dual-use dimension that Anthropic explicitly grapples with: the same capability that helps defenders find and fix weaknesses before deployment could, in principle, be used by adversaries to find weaknesses in deployed systems first. The company's decision to publish this work transparently, rather than keep it internal, reflects a broader strategy of demonstrating both capability and responsible disclosure norms.
More broadly, this research fits into a pattern of Anthropic using its own frontier models to probe scientific and technical frontiers—biology, mathematics, and now cryptography—partly as capability demonstrations and partly as a form of safety research in itself. If AI models are approaching or exceeding human expert performance in cryptanalysis, that has implications for how quickly the security community needs to adapt its assumptions about algorithm longevity, how cryptographic standards bodies vet new proposals, and how much lead time defenders have before AI-discovered vulnerabilities could be weaponized. It also reinforces a theme increasingly common in frontier AI development: as models become more capable of independent technical reasoning, the line between "AI as tool" and "AI as researcher" continues to blur, raising fresh questions about oversight, verification of AI-generated proofs, and how the scientific community should credit and validate machine-assisted discoveries.
Read original article →