Century-old ciphers yield to AI: GPT-6 Astra cracked a 1918 radiogram

In less than a month, models have produced solutions to three ciphers that had been listed as unsolved for decades. They all used the same material: Klaus Schmeh's public list, “The Top 50 unsolved encrypted messages”.

What was cracked. An ADFGVX dispatch from November 27, 1918, 170 characters. The key was found by GPT-6 Astra, an OpenAI model available in ChatGPT. It contained a report on ship movements from those days.

Dates in the plaintext matched the archive. A British cruiser arrived in port on November 24, 1918, and an Allied squadron arrived on the 26th, exactly as the dispatch says.

The gap in the solution. The word TRUPPENVERSCHIEBUNG is documented as a working key only from December 9, 1918, 12 days later than the dispatch itself. The author of the decryption acknowledges the discrepancy and does not explain it.

Independent verification separates two points. The math holds: the archived ciphertext with this key yields coherent German. The key's provenance remains in question, and the corpus custodian has not yet certified the solution.

You can verify it yourself: the cipher-break-verification repository runs independent ADFGVX and Enigma simulators with `python3 check.py` and `python3 -m pytest`. It requires Python 3.10+ and has no external dependencies.

The adjacent case was cleaner. Three days earlier, the same model cracked a Wehrmacht radiogram from July 10, 1941, unsolved for 83 years, in 10 hours. The CryptoCellar archive custodian marked it solved.

Such ciphers used to be broken by enthusiasts, and the work took years. Now it takes hours and minutes: Claude Fable 5.1 solved Thomas Urquhart's 1653 verse cipher in 44 minutes; it had remained unsolved for about 370 years.

The TRUPPENVERSCHIEBUNG key has not yet been certified by the corpus custodian, and until its verdict, the 1918 dispatch is formally still listed as unsolved.