GPT-6 Astra cracks century-old World War I German radio cipher
The AI model has decoded a 1918 German radio message that had eluded human cryptographers, using a key officially recorded as being in use weeks after the transmission.
The artificial intelligence model GPT-6 Astra has successfully decoded a World War I German radio message that had remained unsolved for more than a century. The transmission, sent on 27 November 1918, was part of a list of 50 famous unsolved ciphers maintained by the German science portal Scienceblogs.de. While hundreds of similar German radio messages have been decoded by experts, including codebreaking specialist George Lasry, this particular message had previously resisted solution.
The message was encoded using the ADFGVX method, a German encryption technique that combines a 6x6 grid, known as a Polybius square, with a transposition step. Astra identified the encryption key as “TRUPPENVERSCHIEBUNG”, a German term meaning “troop movement”. This key is documented in J. Rives Childs’s book, “The History and Principles of German Military Ciphers, 1914–1918”. The decoding process involved rearranging the key alphabetically and writing the encrypted message in rows of 19 symbols to reverse the transposition.
Astra’s solution was verified by cross-referencing historical logs of the English cruiser HMS Canterbury. The logs confirm that the ship arrived in Sevastopol on 24 November 1918, followed by an allied squadron on 26 November. These dates align with the content of the decoded message, providing concrete evidence for the model’s accuracy.
The model hypothesised that the cipher remained unsolved due to a discrepancy in the official records. According to the source material, the key “TRUPPENVERSCHIEBUNG” was officially recorded as being in use from 9 December 1918. However, the message in question was transmitted on 27 November 1918, two weeks before the key’s recorded start date. The exact reason for this timeline mismatch remains unknown.
The author of the analysis noted that they were not aware of the message having been decoded previously, describing the result as a minor but significant illustration of the model’s capabilities. The finding highlights the potential of large language models to tackle complex historical puzzles that have stumped human experts for decades.
This development underscores the expanding role of artificial intelligence in fields beyond standard data processing. By applying logical deduction and historical cross-referencing, GPT-6 Astra demonstrated a capacity to solve intricate cryptographic challenges, adding a new data point to the ongoing debate about the utility of AI in specialised research.
