Developer Uses GPT-6 Astra to Decrypt Napoleon's Encrypted Letter, Revealing 1809 Troop Deployments
Developer Carter Church used OpenAI's GPT-6 Astra model for six hours to successfully decrypt a 217-year-old encrypted letter from Napoleon, revealing specific French military deployments in Bavaria and Augsburg. This case demonstrates the practical potential of large language models combined with simulated annealing in historical cryptanalysis.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
According to German tech media outlet t3n.de, Carter Church, an AI engineer at cybersecurity firm SentinelOne, used OpenAI's GPT-6 Astra model to successfully decrypt an encrypted letter sent by Napoleon Bonaparte in 1809. The letter was originally addressed to officer Auguste de Marmont, who was stationed in Dalmatia (present-day Croatia). Due to the imminent war between France and the Habsburg Empire, the content was deliberately encrypted to prevent it from falling into enemy hands. Although Marmont likely knew the decryption method, the key was lost, leaving the letter's contents unread for 217 years.
Church did not start from scratch but utilized 33 characters previously decoded by human cryptographers as a foundation. He ran GPT-6 Astra for six hours using a method called "simulated annealing." This approach generates candidate solutions through trial and error, validating them against frequency statistics common in French literature and Marmont's correspondence. The model eventually identified that some symbols represented entire words rather than single letters, allowing for the full text reconstruction.
The decrypted content reveals that Napoleon informed Marmont of strategic troop deployments, including 40,000 Bavarian soldiers stationed near Munich or Passau and an elite corps located in Augsburg. Church stated he had no background in cryptography and was inspired by another developer, Carter Leffen, who recently used similar methods to decrypt an Enigma message. This event indicates that large language models with strong pattern recognition capabilities can play a significant role when processing historical cryptographic texts with specific linguistic priors.