Anthropic Puts Claude Into Cryptanalysis: Breakthrough or Boundary Test?

Introduction

A recent discussion around Anthropic’s cryptanalysis work raises a difficult question: are large AI models beginning to reach into one of the most demanding areas of computer security? According to the available OSChina summary, cryptographer Matthew Green reviewed Anthropic’s latest results and wrote a commentary on them. The company reportedly used an unreleased Claude Mythos model to perform cryptanalysis, leading to two findings: one involving an attack on the post-quantum signature scheme HAWK, and another improving attacks on reduced-round AES.

That framing requires caution. The source material available here does not include the full original article, the technical paper, parameter choices, proofs, or reproducibility details. So the responsible interpretation is not that AI has broken modern cryptography. Rather, the notable point is that a frontier model appears to have been applied to a highly specialized research task and may have contributed to nontrivial analytical work.

Key points

• The targets matter: HAWK sits in the broader post-quantum cryptography landscape, an area of intense academic and industry scrutiny. AES, meanwhile, is one of the most important symmetric-key algorithms in modern computing, though the reported work concerns reduced-round AES, not the full standardized construction.
• Context is everything: Attacks on reduced-round variants are important for understanding security margins and improving analytical methods. They should not be confused with practical breaks of deployed systems. The same is true for any claim involving a cryptographic scheme: assumptions, parameters, and scope define the actual meaning.
• AI may be moving beyond assistance: Until recently, LLMs were often framed as tools for literature search, coding, or proof assistance. If Anthropic’s results hold up, they suggest that models may help search for attack strategies, combine existing techniques, and generate candidate ideas in a more proactive way.
• Expert review remains non-negotiable: Green’s attention is significant because cryptography is a field where plausibility is not enough. Any claimed result must survive formal reasoning, peer scrutiny, and reproducible validation.

Why it matters

The most important implication is not a sensational claim that “AI can crack encryption.” It is that scientific workflows may be changing. Cryptanalysis depends on deep mathematical intuition, domain knowledge, and familiarity with decades of prior work. If a model can propose useful paths through that space, it could reduce the cost of exploration and help researchers find connections they might otherwise miss.

At the same time, the limits are clear. Cryptographic security is built on precise definitions and rigorous argumentation. A model-generated hypothesis only becomes meaningful when translated into a verifiable proof, algorithm, or experiment. This is especially sensitive in post-quantum cryptography, where attack claims can influence how schemes are evaluated.

For the AI industry, Anthropic’s work is a high-stakes example of AI for Science: applying models not to broad consumer tasks, but to expert-level research problems. Whether the results are ultimately seen as a major breakthrough or a more modest demonstration, they indicate that security researchers may increasingly have to treat AI systems as part of the research environment—both as tools for defense and as accelerators for attack discovery.

Source: OSChina