Anthropic's Claude has solved a cryptographic problem that sat unsolved for 25 years, doing it in under 30 minutes. The problem, originally posed by the National Security Agency, had resisted the efforts of professional mathematicians and cryptographers since it first appeared. Claude worked through it in a single session, producing a solution that researchers have since verified as correct.
What the Problem Was
The specific puzzle falls within a branch of cryptography that deals with number-theoretic structures used to design and analyze encryption systems. Details about the exact nature of the challenge remain limited in public reporting, but sources familiar with the work describe it as a problem where the difficulty lay not in computation but in conceptual framing. Finding the right way to think about the structure was the key, and that is precisely where Claude succeeded. This follows a broader pattern: Claude was credited with solving the 87-year-old Jacobian Conjecture in a separate incident that also surprised the mathematics community.
Key Facts
- The NSA problem had been open for approximately 25 years before Claude attempted it.
- Claude produced a verified solution in under 30 minutes.
- The problem relates to number-theoretic structures used in cryptographic system design.
- Researchers confirmed the solution's correctness after the session.
- The result adds to a growing list of advanced mathematical problems Claude has addressed.
The speed is what has drawn the most attention. Thirty minutes is not a long session by any measure, and the problem was not fed to Claude with any special scaffolding or extended chain-of-thought prompting according to available reports. It was posed in a relatively direct way, and Claude worked through it. That kind of performance on an open problem in applied mathematics is the sort of thing that changes how researchers think about what these systems are actually doing when they engage with formal problems.
The model didn't just find a path to the answer. It reframed the problem in a way that made the answer almost obvious in retrospect.Source familiar with the solution, via The Decoder
Broader Implications for AI and Security Research
Results like this one raise practical questions about how AI systems fit into serious research workflows. Anthropic has been positioning Claude as a tool for advanced reasoning tasks, and episodes like this offer concrete evidence that the positioning reflects something real. At the same time, they prompt harder questions about what it means when an AI system outpaces human experts on problems that were considered genuinely difficult. The security research community has been watching AI capabilities in this space carefully, and the cryptography result will likely accelerate those conversations. Anthropic has previously mapped a year of AI-enabled cyberattacks to the MITRE ATT&CK framework, signaling that the company is already thinking carefully about the dual-use dimensions of capable AI in security contexts.
For now, the episode stands as a data point in an ongoing debate about the practical ceiling of large language model reasoning. Claude's ability to engage with formal mathematical structures has improved with each generation, and Claude's model family has developed increasingly strong performance on tasks that require sustained logical inference rather than pattern matching alone. Whether that represents genuine mathematical understanding or a very sophisticated form of structured search is a question researchers continue to argue over. What is harder to argue with is the output: a correct solution to a problem no human had cracked in 25 years, produced before the half-hour mark.
Anthropic has not issued a formal statement about the cryptography result, and it is not clear whether the company plans to publish the work or submit it for peer review. Given the NSA's involvement in the original problem, there may also be questions about how openly the solution can be discussed. More details are expected to emerge as researchers who have seen the work decide how much to share publicly.
