GPT-6 Astra's WWI Cipher Claim Has a 12-Day Hole in It

A developer publishing under the name Prinz said on Sept. 17 that OpenAI's GPT-6 Astra decoded a German radio message transmitted Nov. 27, 1918, using the ADFGVX cipher, a transposition system that scrambles letter pairs according to a keyword and that the German army used on the Western Front in the war's final year. GPT-6 Astra is OpenAI's flagship reasoning model, positioned by the company as capable of extended, multi-step autonomous work rather than single-turn answers, and it is the same model Prinz credits with independently searching historical archives to check its own translation. The claim, posted to Prinz's Substack, spread quickly on Hacker News and X the same week, framed as a machine succeeding where 108 years of human cryptanalysis had not.

What Astra reportedly did with 170 characters of ciphertext

According to Prinz's account, Astra was given the original German ciphertext, roughly 170 characters long, and told to find the transposition keyword needed to unscramble it. Astra converged on "TRUPPENVERSCHIEBUNG," German for troop movement, and used it to recover a plaintext reading, in translation: "An English cruiser arrived at Sevastopol on the ?4th, an allied squadron follows on the 26th." Astra then cross-referenced that claim against the logs of HMS Canterbury, a Royal Navy light cruiser, and reported that the ship had in fact arrived at Sevastopol on Nov. 24, 1918, with an allied squadron following on Nov. 26, a match Prinz presented as confirmation that the decode was correct rather than a plausible-looking coincidence.

The Neuron corroborates the ship movement, then flags a date that does not line up

A Sept. 18 piece in The Neuron, an AI-focused outlet, checked the historical detail independently and confirmed it: naval history records do show an English cruiser reaching Sevastopol on Nov. 24, 1918, with a squadron following two days later, matching Astra's translation. That corroboration is the strongest part of the claim, since it is a fact about 1918 naval movements that exists in archives independent of anything GPT-6 Astra produced.

But the same piece raised a problem Prinz's post did not address. The Neuron said the keyword TRUPPENVERSCHIEBUNG appears in a historical reference on German military ciphers compiled by J. Rives Childs, an American intelligence officer, and that Childs's account associates that specific keyword with use beginning Dec. 9, 1918, nearly two weeks after the Nov. 27 message Astra says it decoded with that same key. If Childs's dating is right, Astra would have decoded a message using a keyword the German army had not yet started using.

A cryptographer's baseline the claim has not yet been run against

The Neuron pointed to a 2016 academic paper by cryptographer George Lasry and colleagues, which describes decoding 618 of 668 encrypted German texts preserved in Childs's own wartime collection, as the technical baseline any new ADFGVX solve should be measured against. Lasry's paper is a record of which messages from that collection have already been broken and which keywords were in use when. Neither Prinz's post nor The Neuron's response states whether Astra's plaintext was checked against Lasry's list of already-solved messages, which would show whether this is a genuinely new decode or a message cryptographers had already read using conventional methods decades before Astra existed.

What would settle it

Neither account includes an independent party rerunning Astra's process on the original ciphertext image, which is the reproducibility test that would separate a genuine decode from a plausible-looking one. ADFGVX decoding works by testing candidate keywords against expected German phrasing, so a system given enough attempts and a rich enough language model can, in principle, land on a fluent-looking plaintext that happens to fit a real event by chance. Two things would resolve the discrepancy in opposite directions: evidence that Childs's Dec. 9 date is itself incomplete, which a historian of German army signals traffic could check against other intercepts from the same weeks, or an independent cryptographer's confirmation that Astra's proposed key and plaintext are the unique solution consistent with the ciphertext, not one of several equally fluent candidates.

Claim	Status
Historical ship movement (Nov. 24 and 26, 1918)	Corroborated by naval archives, per The Neuron
Keyword TRUPPENVERSCHIEBUNG active by Nov. 27, 1918	Contradicted by Childs's reference, which dates it to Dec. 9
Check against Lasry's 2016 list of already-solved Childs collection messages	Not addressed by either account
Independent reproduction of the decode	Not yet published

OpenAI has not published its own account of the decode, and Prinz's post does not say whether OpenAI reviewed the claim before publication. Until an independent cryptographer reruns the transposition on the original scanned ciphertext and checks it against Lasry's baseline, the strongest available statement is that GPT-6 Astra produced a plaintext that matches a real, independently verifiable naval movement, using a key whose documented usage window does not yet cover the message's date.

The episode arrives weeks after a heap overflow and an SSO misconfiguration reached OpenAI's own internal repositories, and it follows a pattern The Terminal has covered elsewhere this month in AI capability claims: a chatbot's hallucinated intelligence read on a Chinese ship's cargo nearly triggered a military boarding operation before analysts caught the error. Both cases turn on the same question: whether a system's confident, fluent output has been checked against a record that exists independently of the system itself, before anyone acts on it.