'What AI mostly brings is persistence': AIs from Anthropic and OpenAI are cracking ciphers and encryption protocols in record time

• GPT-6 Astra has solved a military code from the 19th century.

• Claude AI has also been investigating a type of encryption based on elliptic curves.

• Both discoveries show how much AI models can do.

AI keeps getting better at many different tasks. Two new reports show how its abilities are growing when it comes to solving secret codes and testing security methods that are widely used to protect information.

First, Japan Today reports that AI engineer Carter Church used OpenAI's GPT-6 Astra model to solve a written code used by one of Napoleon Bonaparte's generals more than 200 years ago. The message was written in 1809 and describes where French troops and troops from allied countries were positioned.

The letter had been sitting unnoticed in an online archive until Church decided to try solving it. Working out the meaning of its symbols, numbers, and letters by hand could have taken weeks. The AI solved it in just six hours.

"What AI mostly brings is persistence," Church told AFP. He added that he felt "an unbelievable sense of satisfaction and excitement" when the answer appeared. Patent engineer Satoshi Tomokiyo checked the result. Tomokiyo keeps an online list of unsolved codes that amateur codebreakers can try to solve.

Curves and keys

The second AI discovery is about how cryptographic keys are created. These keys are used by software and hardware to encrypt information and keep it safe. One way to create these keys uses shapes called elliptic curves. A "seed" is a starting value used to create a curve.

According to Cyber Security News, an internal version of Anthropic's Claude AI has worked out where five elliptic curve seeds created by the US National Security Agency (NSA) came from. It did this in less than 30 minutes. The seeds were published more than 25 years ago, and people have long wondered whether they contained hidden weaknesses that the NSA could use to secretly get around the security they were meant to provide.

Claude's finding was that they did not. The seeds had no weaknesses and contained no hidden jokes. As Anthropic explains, most researchers already trusted these curves, so no action is needed. Still, the result is a "fun closing chapter" that provides another example of what AI can do.

Claude also made a pointed comment about human code experts. According to the logs showing its internal reasoning, the AI asked, "Is it suspicious that such an obvious thing wasn't discovered in 25 years?" as it worked toward the correct answer. These models are not only doing well at the difficult mathematics involved, but also making fun of people along the way.