Claude solves in 30 minutes NSA mystery that remained unsolved for 25 years

Anthropic's AI model Claude solved a mystery security experts had been struggling with for over 25 years in less than half an hour. These are mysterious numbers that the US intelligence service NSA used to develop an important security standard.

For years, it was suspected that the secret service had built in a hidden back door to be able to crack encrypted data. Now it turns out that the numbers probably have a much simpler origin.

The story begins in the late 1990s, when the US intelligence agency NSA provided mathematical data for a new security standard. It was adopted by the American standards institute NIST in 2000. The technology used helps, among other things, to secure internet traffic, digital information and financial systems.

But there was something strange about the security method. The NSA had used certain numbers without explaining how they had been chosen. This caused distrust among security experts. They wondered if the secret service had deliberately selected numbers that made security easier to crack.

Those suspicions were later reinforced by Edward Snowden's revelations. This revealed that the NSA had been involved in attempts to weaken cryptographic security. The American NIST eventually even deleted another, controversial method from its recommendations.

Researchers spent years trying to figure out the origins of the mysterious numbers. They examined old documents, approached government agencies, and searched for clues from NSA employees involved. But there was no convincing explanation.

That changed unexpectedly this month. An Anthropic employee instructed an internal research model of Claude to investigate the riddle. Without giving previous theories about the origin of the numbers, he let the AI model itself look for an explanation.

Within half an hour, Claude found a surprising answer. The numbers were found to be reconstructable withlrand48, an old computer program that generates random looking numbers. Using dates from 1997 and 1998 as a starting point, the model was able to mimic the original calculations.

With the commonly used security method P-256, Claude found an exact match with the published number for the 1,842nd candidate. Four other mysterious numbers could also be traced in the same way.

The discovery even surprised the AI model. In his recorded reasoning, Claude initially wondered whether the outcome could be real, because according to his knowledge the mystery was still unsolved.

When the calculations proved to be correct, the model enthusiastically wrote: "Wow, this is a real discovery."

The outcome makes it less likely that the NSA deliberately selected the numbers examined to create a hidden vulnerability. Instead, they appear to come from an existing computer program, the calculations of which can now be repeated by anyone.

This is important, because the lack of an explanation caused mistrust for years. Security experts were able to see the numbers, but could not verify why the NSA had chosen exactly this one.

However, not all questions have been answered yet. For example, researchers don't yet know why the original program skipped some useful numbers. It is also not entirely clear how certain other elements of the security methods were established.

According to Anthropic, users of the security standards involved do not need to change anything. The discovery did not reveal a new vulnerability, but above all provided clarity about a question that had remained unanswered for years.

There was also a hefty reward for solving the mystery. In 2023, cryptography researchers offered $12,288 for the one who could figure out where the mysterious NSA numbers came from.

The reward went up to $36,864 if the prize money was donated to a charity. Now that Claude has solved the riddle, that full amount goes to the Internet Security Research Group.

This nonprofit is behind Let's Encrypt, a service that helps websites secure their connections with free security certificates.

This gives the discovery a special follow-up. An AI model not only solved a puzzle that remained unsolved for a quarter of a century in less than thirty minutes, but also generated tens of thousands of dollars for the security of the Internet.