Claude, Anthropic's AI assistant, has solved a cryptographic problem that the National Security Agency left open for 25 years, completing the work in under 30 minutes. The puzzle, which had challenged mathematicians and cryptographers for decades, yielded to Claude's analysis in a session that observers are now scrutinizing for what it may signal about AI capability in formal reasoning and number theory.
What the Problem Was
The specific challenge involves a number-theory puzzle tied to NSA research from the late 1990s. For full technical background on what Claude solved and how the session unfolded, see our earlier coverage of Claude cracking the 25-year NSA number mystery. In short, the problem required identifying a structured mathematical relationship that had resisted straightforward computational brute force and had not succumbed to the approaches tried by human researchers over the intervening years.
Key Facts
- The NSA puzzle had been open for approximately 25 years
- Claude completed the solution in under 30 minutes in a single session
- The problem falls within number theory and cryptographic analysis
- The result was reported by Ynetnews and corroborated by independent coverage
- No specialized cryptography plugin or tool was used; Claude worked through the problem using reasoning alone
What makes the episode notable is not just the speed but the method. Claude did not rely on exhaustive search. Instead, it appears to have identified structural properties of the problem that allowed a more direct path to a solution. That kind of reasoning, sometimes called mathematical insight, is typically associated with experienced human researchers rather than automated systems. Whether this represents a genuine step forward in AI mathematical capability or a well-matched problem is a question researchers are now debating.
The fact that it got there in 30 minutes, on a problem that stumped people for 25 years, is going to make a lot of cryptographers pay close attention.Independent security researcher, as reported by Ynetnews
Context and Implications
This result arrives as Anthropic has been pushing its models into more technically demanding domains. The company recently launched Claude Science, a product aimed at research-heavy industries including pharmaceuticals, and has been expanding the contexts in which Claude is tested against hard scientific and mathematical problems. A 25-year cryptographic puzzle falling in under half an hour fits a broader pattern of Claude performing beyond expectations in structured reasoning tasks.
The security dimension is worth considering carefully. Cryptographic research has direct applications in how data is protected at scale. Earlier this year, Anthropic published work mapping AI-enabled cyber activity to established threat frameworks, a project that reflects the company's awareness that capable AI models can be relevant to both offense and defense in security contexts. A model that can solve open cryptographic problems quickly is a tool that security professionals and policymakers will want to understand thoroughly.
It is too early to say whether this single result generalizes. One solved puzzle, however old, does not establish that Claude can routinely defeat open problems in mathematics or cryptography. What it does suggest is that for certain classes of structured, well-defined problems, large language models may be further along than the research community had assumed. That question will drive a lot of follow-up testing over the coming months.
For readers tracking how Claude's capabilities are evolving across different domains, this episode adds a concrete data point to what has otherwise been a year of incremental announcements. The cryptography result is specific, verifiable, and significant enough that it will likely inform how researchers design future benchmarks for AI mathematical reasoning.
