An AI model cracked a two-century-old French military cipher in about six hours, using nothing but a single photo of the code.
Carter Church, an engineer at cybersecurity company SentinelOne, fed GPT-6 Astra a single photo of a 1809 cipher and asked it to solve it. The message came from the headquarters of Eugene de Beauharnais, Napoleon's stepson and Viceroy of Italy, and relayed Napoleon's orders to French General Auguste de Marmont. It had sat unread on Cryptiana's Unsolved Historical Ciphers list for decades, its 24 rows of numbers, letters, and invented symbols untouched. Astra worked through the puzzle with a partial key nobody had previously connected to the letter, then closed the gap with its own simulated-annealing solver, and Cryptiana curator Satoshi Tomokiyo has since marked it solved.
The interesting part isn't really the codebreaking - 19th-century field ciphers are well within reach of modern cryptanalysis with or without AI. What's notable is that Astra ran the whole pipeline - reading the handwriting, identifying the cipher structure, and solving it - from one prompt and one image, with no human stitching the steps together. That kind of end-to-end multi-modal reasoning, not an old cipher finally giving up its secrets, is the real story here.
The decoded text also fills a gap in Napoleon's 1865 memoir, which breaks off mid-order with Marmont told not to be slowed by "a handful of..." Church says the cipher supplies the missing words, though the published phrase, "a gathering of rabble," reads as more paraphrase than seamless continuation - the solution package is public if you want to judge for yourself.