Exhibitor login
AI Insider 17 September 2026

Carter Leffen cracks Enigma with GPT-6 Astra

Carter Leffen cracks Enigma with GPT-6 Astra

Carter Leffen, product development coach at Bloomberg LP, has managed to crack a radio message of 82 characters from 1941 that was encrypted with the Enigma machine. This news was announced on September 17, 2026, in a detailed case study. The message pertains to a German soldier reporting his location in Rosenow and requesting marching orders, a message that had escaped 83 years of encryption.

The breakthrough was achieved through the use of OpenAI's GPT-6 Astra, which unraveled the core of the issue in about ten hours. By utilizing a previously deciphered message from the same day and a weakness of the Enigma machine, Leffen was able to find the correct key. The Enigma machine does not encrypt a letter as itself, which allowed for quick exclusion of incorrect positions. Leffen and his AI agents used the place name Rosenow as a search term to uncover the encrypted message.

The AI variant "GPT-6 Astra Extra High" performed a series of complex tasks, including setting up a simulation, cryptanalytic coding, and conducting parallel trials. The deciphered text includes instructions such as “Sofort Funkantwort” and “Angabe des Marschweges,” while also showing some spelling mistakes that indicate real transmission errors. Leffen sees these mistakes as confirmation of the authenticity of the decryption. This approach utilizes existing cryptanalytic insights and emphasizes the collaboration between human expertise and AI technology.

When Leffen stated that building the website to share the problem and solution required more effort than the cracking itself, he highlighted the importance of documentation and reproducibility in such cases. Although the code and simulators are openly available for download, it remains to be seen whether independent experts can verify the solution. This project demonstrates how modern AI tools coalesce with classical cryptography, but there is still a need for external evaluation.

Read the full article from AI Insider.