Model GPT-6 Astra deciphered a historical code from the time of World War I

Researchers applied the neural network GPT-6 Astra to crack the historical ADFGVX cipher used by the German army during World War I. The model successfully restored the content of an encrypted radio message that had remained undeciphered for over a century.

In the experiment, the system handled the task in 1 hour, receiving a score of 8 points on the internal complexity metric. The final key for decryption was the phrase «TRUPPENVERSCHIEBUNG». The decrypted text contains data on the arrival of an English cruiser in Sevastopol, dated November 24, 1918.

The ADFGVX cipher is a combination of a transposition cipher and a substitution cipher using the Polybius square. Previously, similar tasks required significant computational power and the involvement of cryptanalysts, but the use of modern generative architectures allows for the automation of the key search process in large variant spaces.

The results of GPT-6 Astra demonstrate the ability of models to solve specialized tasks in the field of historical cryptography. The successful decryption confirms the effectiveness of reasoning algorithms when working with unstructured data and historical archives, where there are no direct instructions for cracking methods.