§ Claim under review · Capability
"OpenAI's GPT-6 Astra model helped crack a Napoleonic military cipher that had remained unread for 217 years, decoding a March 1809 letter sent to French General Auguste de Marmont shortly before Austria invaded Bavaria and triggering the War of the Fifth Coalition."
Verdict
Mostly accurate
Confidence
MediumSummary
This one largely checks out. GPT-6 Astra is a real OpenAI model released in September 2026, so the suspicion that it was invented is wrong. An engineer at the security firm SentinelOne, Carter Church, published a full reading of a coded 1809 letter to French General Marmont that had been listed for years as an unsolved historical cipher, and he credits the model with transcribing the symbols and rebuilding the key over about six hours. Named historians at King's College London and the University of Portsmouth have called the work credible and impressive. Two things the post leaves out are worth knowing. About a third of the cipher's letter values had already been published by an earlier researcher and were used as the starting anchors, and because Napoleon's separate instructions describing what the letter should say still survive, one historian said the decoded text does not actually tell historians anything substantially new. The solution is self-published rather than peer reviewed, and no independent expert verification of the full key has been confirmed so far.
The readings
key figures from the evidenceyears the Marmont cipher remained unsolved before decoding
Why this verdict
Evidence
GPT-6 Astra is a real, publicly released OpenAI model. OpenAI's own channels announced it on September 3, 2026 with rollout to ChatGPT Plus, Pro, Business and Enterprise users and to the API, and it has a model card on Amazon Bedrock. The "GPT-6 Astra is not a real product" suspicion recorded in the intake note is incorrect as of today.
The cipher work is also real and traceable to a named individual with a published, self-contained write-up. Carter Church, an AI engineer at the cybersecurity firm SentinelOne, published "Breaking the Marmont Cipher, 1809," presenting a full reading of a ciphered letter addressed to General Marmont. His account states the letter had been catalogued as unsolved on Satoshi Tomokiyo's Cryptiana list of unsolved historical ciphers, under the entry title "Encoded Letter to Marshal Marmont (1807)," and that the only public image is a plate printed in a 1969 French army journal article by J. Vilcoq, carrying what he argues is the wrong date.
Per the write-up: the ciphertext contains 1,300 cipher units drawn from 155 distinct signs, it is homophonic, and a partial table compiled by French cryptology historian Daniel Tant and hosted by ARCSI already documented 33 letter values, covering 435 of the 1,300 units, roughly a third. The remaining roughly two thirds were reconstructed, with the model building a simulated annealing solver scored on resemblance to French. Church says the model worked from a single low-resolution scan, about 1,202 by 1,836 pixels, and ran about six hours. He redates the letter to late March 1809 and identifies it as originating from the headquarters of Eugène de Beauharnais, Viceroy of Italy, Napoleon's stepson, sent on Napoleon's instruction of March 16 to brief Marmont, then commanding in Dalmatia, on the disposition of French and allied armies.
Two named academic historians commented to Live Science. Zack White of the University of Portsmouth described it as a credible solution to a historical mystery. Michael Rowe of King's College London called the findings impressive and explained why confidence is reasonable: Napoleon's original correspondence to Eugène survives and specifies what the ciphered message should contain, so the decoded text can be checked against a known expectation. Rowe described this as a kind of Rosetta Stone case, and added that for the same reason the result does not tell historians anything substantially new.
Findings
✓ What's accurate 7
- GPT-6 Astra is a real OpenAI model, publicly announced on OpenAI's own channels on September 3, 2026 and available in ChatGPT tiers, the API, and via AWS. The intake note's doubt on this point is not supported.
- A ciphered letter addressed to Marmont was listed as unsolved on a standard reference list of unsolved historical ciphers, known publicly only from a 1969 French army journal reproduction catalogued with a different date.
- A named engineer at SentinelOne, Carter Church, published a full proposed reading of that letter, attributing the transcription and cipher reconstruction to GPT-6 Astra working from the single scan over roughly six hours.
- The figures in the post match the write-up: 1,300 cipher units, 155 distinct signs, 33 letter values previously published by Daniel Tant.
- The letter is dated by the solver to late March 1809 and is placed roughly two weeks before Austria invaded Bavaria in April 1809, the action that began the War of the Fifth Coalition. Marmont was at that time a general commanding in Dalmatia; he was made a Marshal later in 1809, so the post's use of "General" is correct for the date in question.
- The decoded content as described, positions and strength of French and allied armies, Austrian movements, and reassurance to Marmont, matches what the write-up and the news coverage report.
- The claim's verb "helped" is the right strength. It does not assert the model acted alone.
≈ What's misleading 4
- The post presents the cipher as unreadable with "only 33 letter values previously documented," which is accurate, but omits that those 33 published values covered about a third of the ciphertext and served as the anchors from which the rest was reconstructed. The starting point was a partially solved cipher, not a blank one. This matters for how hard the problem was and for how far the result generalizes to ciphers with no partial key.
- The post does not convey that the historical content was largely predictable in advance. Napoleon's surviving instruction to Eugène specifies what the ciphered letter was supposed to contain, which is why historians consider the reading trustworthy, and it is also why one of those historians said the decoded text does not reveal substantially new information. The post's framing, "revealing the contents," reads as a historical discovery when the expert assessment is closer to a confirmation of an expected text.
- The solution is a self-published proposal by its author, with some signs still ambiguous by his own account because they occur only once. No peer review, no confirmed independent cryptologic replication. The post presents it as a settled crack.
- **Demo to product conflation, mild:** the post is headlined as an AI capability event with the model as the actor. The described workflow was selected, prompted, supervised and verified by a human engineer who chose this specific problem to showcase the model, and who has said publicly that verification was the hardest part and cannot be automated. The claim's own wording "helped" is faithful; the surrounding presentation as an AI news item about what the model did is where the drift occurs.
? What's uncertain 4
- Whether the reading is correct in full. It is checkable in principle because the key and a decoder were released, but no confirmed independent cryptologic verification was found. One social-platform summary asserts that the maintainer of the unsolved-cipher list verified it; that assertion could not be corroborated.
- Whether the model's role was as autonomous as described. The account of what the model did unaided, including the six-hour runtime and the claim that it located the Tant table through its own web research, rests entirely on the author's self-report. No log, transcript, or third-party observation was found.
- Whether the redating from 1807 to March 1809 is accepted by historians of the period. The two historians quoted addressed the plausibility of the decipherment rather than endorsing the redating specifically.
- Whether the contamination control described by the author, rerunning the solver with Napoleonic texts removed from the scoring corpus, was independently checked.
Sources
6 of 8 linked to recordsCarter Church, "Breaking the Marmont Cipher, 1809," carter.church/writeups/the-letter-to-marmont/
OpenAI, "GPT-6 Astra" launch page and OpenAI's official account announcing the rollout on September 3, 2026
Amazon Bedrock model card for GPT-6 Astra
Live Science, "A kind of Rosetta Stone: AI decodes 217-year-old secret letter ordered by Napoleon," 2 October 2026, including direct comment from two named academic historians
Euronews, "AI decodes 217-year-old Napoleonic letter in 6 hours," 2 October 2026
Calcalist, "SentinelOne AI researcher uses GPT-6 Astra to crack Napoleon cipher after 217 years"
Satoshi Tomokiyo, Cryptiana "Unsolved Historical Ciphers," entry "Encoded Letter to Marshal Marmont (1807)"
Various aggregators and syndications (Yahoo Tech, AOL, ground.news, All That's Interesting, kucoin, 36kr, vgtimes, azernews)