Napoleon Kept This Secret for 217 Years, AI Broke It in Six Hours

In 1809, a coded letter went out to one of Napoleon's generals. It was written by the emperor's stepson, on the emperor's direct orders. About 1,300 characters of strange symbols. And then the key to reading it vanished. For 217 years, nobody could say what it said.

Now an AI model has read it. In about six hours.

So what did the AI actually do?

It was OpenAI's GPT-6 Astra, and an engineer at a security company pointed it at a single scanned image of the letter. Then it did the lot. It transcribed the image, worked out that this was a homophonic substitution cipher (one letter can hide behind several different symbols), wrote its own code, and cross checked historical records along the way. The letter turned out to be a briefing for General Marmont: where Bavarian, Polish and French forces were, what the Russians were up to, and Napoleon's view that Austria was heading for defeat. He even told the general not to be rattled by a few troops or a gathering of rabble.

But why did two centuries of human experts miss it, when an AI got there in an afternoon?

Here is where it gets fun. It was not one magic leap. A related cipher table had been sitting in the archives the whole time. Reportedly it was catalogued under the name of a different general, a one letter slip that looks like a plain naming mistake. So for a very long time nobody had a reason to connect that table to this letter. The AI tried applying it anyway. And it worked. One historian who looked at the case describes it that way, so treat the filing mistake as the likely explanation rather than something proven.

But the filing error is only half the story. The scan was a big plate with 175 hand drawn marks to sort out, and a nineteenth century clerk drew the same sign slightly differently each time. Someone, or something, had to work out which squiggles were the same squiggle before the codebreaking could even begin. That is tedious, error prone, mind numbing work. Exactly where a machine that never gets bored shines.

Did the AI just remember the answer from its training? The engineer checked. According to his own write up, he reran the whole thing with anything about Napoleon and Marmont stripped out of what the model could draw on, and it came back with the same reading. The person who keeps a well known database of unsolved historical ciphers has since marked this one as solved.

So no, not a flash of genius. A missed clue, a lot of patient pattern matching, and a machine that never got tired. Which might be the most interesting kind of breakthrough there is.