GPT-6 Astra deciphers a 217-year-old Napoleonic-era encrypted letter
A person with no codebreaking experience gave Astra an image of an encrypted letter to a Napoleonic general, and it spent about six hours transcribing 1,300 signs and writing code to crack the cipher, with the result confirmed by the historical cipher archive's maintainer.
just insane. you can use AI to crack unsolved historical mysteries as a hobby now. here’s how to find a mystery and try it yourself: you don’t need any experience in historical research or codebreaking, and there are actual lists with hundreds of real mysteries to choose from. for example, carter church had never broken a cipher before. he gave GPT-6 Astra one image of an encrypted letter to one of napoleon’s generals and a goal: decipher it. Astra spent about 6 hours transcribing 1,300 signs, finding a partial key, and writing code to crack the rest. the result: a readable 217-year-old war briefing, with the solution confirmed by the historical cipher archive’s maintainer. here’s how to try it yourself: 1. pick a real mystery. start with one of these lists (links in the reply): > Cryptiana’s unsolved historical ciphers > Daniel Bourdeau’s unsolved catalogue > DECODE database > Klaus Schmeh’s encrypted book list choose one with a readable scan, and check the latest notes to make sure nobody's solved it yet 2. give the AI the evidence and the goal. the only 2 models i’d use for this are GPT-6 Astra in Codex or Opus 5.5 in Claude Code. upload a clear image of the encrypted document, along with the source link. then ask it to decipher the message. tell it to research existing clues and previous attempts, then write and run code to test possible solutions. 3. make it show its work. if it finds a solution, ask for the decoding key, code that reproduces the reading, and anything still uncertain. then you have something concrete for the archive maintainer or a specialist to check. 4. email the list’s maintainer. for Cryptiana, that’s Satoshi Tomokiyo: verlat@hotmail.com. send a short summary with the original document link, your proposed reading, and the decoding key and code from step 3. mention that you used AI, flag anything still uncertain, and ask whether they'd be willing to review the solution. if they confirm your solution, congrats. you just cracked a historical mystery! it sounds almost too good to be true, but some of these ciphers are actually relatively straightforward for a specialist. there are just a lot of them scattered across digitized archives, and specialists only have so much bandwidth. but now that the frontier models are good enough, a curious hobbyist can pick one, give an AI agent the evidence, and put it to work. good luck, and go dig up some centuries-old secrets :)
This post quotes another post: “GPT-6 Astra decrypts a 217-year-old cipher to one of Napoleon's generals”