Carter Church spent six hours and a GPT-6 Astra session to read a letter no one had cracked since 1809. The sender was Napoleon's own stepson.
Carter Church, a staff AI engineer at the cybersecurity firm SentinelOne, fed a single low-resolution scan into OpenAI's GPT-6 Astra and walked away six hours later with a fully translated letter that had sat unread since March 1809. The document was dispatched from the headquarters of Eugène de Beauharnais, Napoleon's stepson and the viceroy of Italy, to General Auguste de Marmont, as Austria moved toward war with France. It ran to roughly 1,300 cipher units built from 155 distinct symbols, a homophonic substitution cipher designed so that common letters could hide behind several different signs at once, which is exactly the kind of trick that makes basic frequency analysis useless.
Nobody had the key. Historian Daniel Tant had previously worked out 33 of the symbols, covering maybe a third of the message, and that's where the trail went cold for decades. Astra started from Tant's partial progress, transcribed the remaining symbols straight off the scanned image, and then did something no earlier attempt had managed: it wrote its own code. According to reporting from Euronews, the model built a solver based on simulated annealing, a trial-and-error search technique that tests thousands of possible letter and word assignments against the remaining unknown symbols until a coherent text falls out.
It worked. And along the way, Astra caught something the historical record had gotten wrong. Calcalist reported that Church used the decoded content, troop numbers, officers' titles, unit positions, to re-date the letter from 1807 to March 1809, just weeks before Austrian forces actually launched their offensive in April. That's not pattern-matching. That's cross-referencing a decrypted military briefing against known troop movements and catching a dating error historians had carried for over a century.
Once translated, the text turns out to be exactly what you'd expect from a command briefing written days before a war: French, Bavarian and Polish troop positions, laid out unit by unit, meant to brace one general for an isolated and dangerous position. Napoleon's own words come through in the message, urging Marmont not to be intimidated by Austria's buildup. It's a small, human detail buried inside two centuries of silence, an emperor talking a subordinate off the ledge before the shooting starts.
OpenAI Says GPT-6 Astra Marks the Start of the AGI Era OpenAI launched GPT-6 Astra on September 3, 2026, with president Greg Brockman declaring "Welcome to the AGI era." The same model is the first from OpenAI to hit the "Critical" cybersecurity threshold in its Preparedness Framework, meaning it can find and exploit unknown software vulnerabilities on its own. - openai GPT-6 astra artificial general intelligence - when did openai release GPT-6 astra model
Church didn't just announce the result and move on. He published his full research record on September 18, including the transcription, the cipher tables, and the scripts, so that other cryptologists and historians can reproduce the work line by line. That's the part worth sitting with. A closed claim from an AI lab is marketing. An open, reproducible decryption that any historian can check against the archive is something else entirely.
This matters beyond the history books. GPT-6 Astra wasn't handed a clean dataset or a known answer key. It was given one blurry photograph of a 217-year-old document and asked to do four genuinely different jobs in sequence: read handwriting and symbols off a bad scan, diagnose the type of cipher it was looking at, write original code to attack that specific cipher, and then cross-check its own output against real historical records to fix a date nobody else had caught. Frontier labs spend most of their public demos on math olympiad problems and coding benchmarks. This is a case where the model's reasoning and tool-use showed up on a problem nobody built a benchmark for, because nobody expected it to be solvable at all.
Cryptologists have spent real careers on exactly this kind of archive. Historical cipher-breaking is slow, often thankless work, and the backlog of unread 18th and 19th century military correspondence sitting in European archives is large. If a model can do in an afternoon what took professional cryptanalysts nowhere over 200 years, the honest question isn't whether AI is impressive. It's how many other sealed letters are sitting in a drawer somewhere, waiting for someone to point a scanner and a frontier model at them.
Also read: Hawley and Murphy bill would send AI executives to prison over rogue agent hacks • Bittensor subnets are generating real revenue, but the cheap prices hide a subsidy • California tech CEO charged in $300 million scheme to smuggle Nvidia chips to China
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.
OpenAI Changed GPT-6 Astra's Benchmark Numbers Days After Its Launch Fortune reported that OpenAI quietly revised several GPT-6 Astra benchmark figures after its September 3 launch, including cutting its hallucination rate in half before later reverting it, and boosting a cybersecurity score using a reasoning tier that isn't commercially available. The changes mostly flattered Astra, though some of Anthropic's... - OpenAI changed GPT-6 Astra benchmark numbers after launch - how OpenAI modified benchmark results for GPT-6 Astra