The OpenAI GPT-6 Astra model cracked a 217-year-old encrypted letter in just six hours. The OpenAI GPT-6 Astra model cracked a 217-year-old encrypted letter in just six hours, according to Carter Church, an AI engineer at SentinelOne who reported on the decryption.
This document contains 24 lines of encrypted text, totaling around 1,300 code units. These units employed 155 different symbols, making the decryption process quite intricate. The AI model worked from a low-resolution image of the letter, published in a 1969 French military history journal, which added another layer of complexity to the task.
The process started with the AI recognizing and transcribing the various handwritten character variants. After conducting online research, it discovered a partial key from cryptologist Daniel Tant to help evaluate 33 symbols, which only clarified about a third of the text. For the remaining sections, the model created its own search algorithm based on patterns in the French language.
The letter was written by Eugène de Beauharnais, Napoleon’s stepson, and addressed to General Auguste de Marmont, who was stationed in Dalmatia at the time, isolated from other French forces. It informed Marmont about the positions, commanders, and troop numbers of French and allied forces, and instructed him not to be deterred by smaller enemy forces or gathered crowds. The encryption protected that military information.
Experts like Michael Rowe from King’s College London suggest that historians could have cracked the code with enough time and effort. However, he points out that the letter does not reveal a completely new military strategy, as it reiterates known orders from Napoleon. Zack White from the University of Portsmouth noted that the letter’s inflated troop numbers might raise intriguing questions about Napoleon’s assessment of his situation.
Carter Church, an AI engineer at SentinelOne, published the program and data used for verification, though his own write-up notes uncertainty about the values of five symbols in the cipher.
