AI Model Reportedly Cracks an Unsolved 1918 German ADFGVX Cipher

A blog post claims that an AI system, referred to as “GPT-6 Astra,” decoded one of the roughly 50 famous unsolved ciphers cataloged by Scienceblogs.de — specifically a German WWI radio message sent on November 27, 1918, and encrypted with the ADFGVX method. ADFGVX maps each character to a two-symbol coordinate in a keyed 6x6 grid and then transposes the result using a keyword, so recovering a message requires finding the right key and column ordering. Hundreds of these intercepts have already been broken, notably by cryptanalyst George Lasry, but more than a dozen had resisted every attempt, including this one.

The model reportedly identified the key “TRUPPENVERSCHIEBUNG” — documented in J. Rives Childs’s history of German military ciphers — and produced a plaintext reporting that an English cruiser had arrived at Sevastopol and that an Allied squadron would follow. It then checked its own work against historical records, matching the decode to HMS Canterbury’s arrival at Sevastopol on November 24, 1918, and a subsequent Allied squadron on the 26th. Its explanation for why the message stayed unsolved: that key was believed to have entered use only on December 9, whereas this transmission predates it, so prior solvers likely never tested it.

The significance, if it holds up, is a concrete example of an LLM performing genuine historical cryptanalysis — searching a key space, applying a non-trivial transposition procedure, and corroborating the output against primary sources. The caveat is that this is a single, self-reported result from the author’s own account, not an independently verified break, and the discrepancy in the key’s supposed introduction date remains unexplained. It reads as an intriguing demonstration that warrants confirmation from the codebreaking community rather than a settled finding.