GPT‑6 Astra Deciphers a World War I German Radio Cipher

The recent decryption of a World War I German radio transmission highlights how modern language models can assist classic cryptanalysis.

The message, sent on 27 November 1918, had remained on the “unsolved ciphers” list maintained by the German portal Scienceblogs.de.

That list catalogs fifty historic puzzles, ranging from serial‑killer cryptograms to the Voynich manuscript.

The encrypted traffic employed the ADFGVX method, a polygraphic cipher introduced by the German Army in 1918.

ADFGVX uses a six‑letter square (A, D, F, G, V, X) for both rows and columns to encode each plaintext character as a pair of symbols.

For example, with the key word “HOUSE”, the square maps “AA” to H, “AD” to O, and “DA” to B.

The key word determines the permutation of the alphabet and digits within the square, making each key produce a distinct table.

Historical records show that German operators used a handful of known keys, allowing cryptanalyst George Lasry and others to decode hundreds of messages.

However, more than a dozen transmissions, including the November 27 broadcast, resisted all manual and automated attempts.

Understanding the ADFGVX Cipher

The cipher was adopted for high‑security radio traffic because its double‑symbol encoding increased the difficulty of frequency analysis.

Nevertheless, once the key word is known, the square can be reconstructed and the ciphertext reduced to a simple substitution.

GPT‑6 Astra’s Breakthrough

In September 2026, the AI model GPT‑6 Astra announced it had solved this particular cipher.

Astra reported the plaintext as “EIN ENGLISCHER KREUZER EINLIEG X SEWASTOPOL X S4STEN X EIN GESCHWADER DER X ALLIIERTEN FOLGT 26STEN X”.

Translated, the text reads “AN ENGLISH CRUISER ARRIVED AT SEVASTOPOL ON THE ?4TH AND ALLIED SQUADRON FOLLOWS ON THE 26TH”.

The model selected “TRUPPENVERSCHIEBUNG” as the encryption word, following the procedure described on pages 214‑215 of J. Rives Childs’s 1914 reference on German military ciphers.

To use the key, the letters of “TRUPPENVERSCHIEBUNG” are first alphabetically reordered, then placed horizontally above the ciphertext arranged in rows of nineteen symbols.

This arrangement creates eighteen columns of nine symbols and one column of eight symbols, allowing the model to map each symbol pair back to a plaintext letter.

Astra’s internal calculation identified the first pair “AV” as corresponding to the letter E, beginning the decoded message.

The model repeated this systematic lookup until the entire transmission was reconstructed.

Astra hypothesised that the reason the cipher had remained unsolved was a mismatch between the key’s official start date (9 December 1918) and the message’s earlier transmission date.

To verify the result, Astra cross‑checked naval logs and found that the British cruiser HMS Canterbury indeed docked in Sevastopol on 24 November 1918.

The logs also recorded an allied squadron arriving on 26 November, matching the decoded reference to a “26STEN” squadron.

The author of the original post noted, “I am not aware of this particular message having ever been decoded before, so sharing it here as a minor (but I think really cool) result and illustration of the capabilities of this model.”

This acknowledgment underscores that the decryption was not previously documented in academic or hobbyist circles.

The breakthrough demonstrates that large language models can automate the labor‑intensive steps of classical cipher analysis, such as key permutation and columnar transposition.

It also suggests a new workflow where historians supply known cipher specifications and AI models generate candidate plaintexts for expert review.

Nonetheless, the process still requires human validation, as shown by Astra’s reliance on historical ship logs to confirm the meaning of ambiguous characters.

Future research may extend this approach to other unsolved ciphers on the Scienceblogs.de list, potentially unlocking further wartime communications.

The episode illustrates a concrete synergy between AI capabilities and traditional cryptographic scholarship.

Why This Matters