Particle.news
Download on the App Store

AI Decodes 217-Year-Old Napoleonic Cipher

GPT-6 Astra used a partial 33-symbol key and computational solvers to recover a routine 1809 military briefing, demonstrating AI can aid but not guarantee historical decipherments.

Overview

  • Carter Church, an AI engineer at SentinelOne, used OpenAI's GPT-6 Astra in early October 2026 to reconstruct the Marmont cipher that had gone unread since 1809.
  • The cipher spans about 1,300 cipher units and roughly 155 unique signs and is a homophonic substitution system in which some symbols stand for whole French words.
  • The workflow combined Astra's image transcription with a partial key compiled by historian Daniel Tant, then ran computational cryptanalysis similar to simulated annealing to recover remaining symbols.
  • Independent reviewer Satoshi Tomokiyo of Cryptiana and public release of Church's key, transcription and code improved reproducibility though five symbol readings remain uncertain.
  • Historians say the decoded text is a routine troop briefing that aligns with existing records, and they warn the result depended on favorable prior materials and careful human verification rather than proving universal AI success.