OpenAI's GPT-6 Astra Cracks 83-Year-Old Enigma Message in Ten Hours

A Bloomberg developer says he used OpenAI's GPT-6 Astra to decrypt a Wehrmacht Enigma transmission that had gone unread since 1941, solving it in roughly ten hours. The message runs 82 characters and, according to the claimed plaintext, has a German soldier asking about his march route. If the reading holds up, it would be the first public case of a frontier language model breaking a real wartime ciphertext rather than a constructed puzzle.

The claim matters because Enigma is not a single lock but a family of them. Operators set daily rotor orders, ring positions and plugboard pairings, and the same ciphertext can yield different plaintexts under different settings. Historic breaks, from Poland's Biuro Szyfrów in 1932 to Alan Turing's work at Bletchley Park, leaned on stolen key sheets, cribs and specialized bombes. A general-purpose model has none of that hardware or captured material behind it.

Details of the run remain thin. The developer has not published the prompt chain, the candidate keys tested or the model configuration, so it is unclear whether Astra searched the key space itself, leaned on a crib, or was guided step by step. That gap is the main reason cryptographers are withholding judgment. Ten hours is impressive against a problem that stood for 83 years, but it says little until the method is reproducible.

Independent review is the next gate. The standard check is straightforward: publish the ciphertext and the settings, then let a second party re-run the decryption and confirm the German reads cleanly. Enigma messages are short and formulaic, which makes false positives possible, especially when a model is encouraged to produce fluent output. A plausible sentence about a march route is not the same as a verified one.

There is also a category question. Classical ciphers are a weak benchmark for modern language models, because the search space is small by today's standards and the underlying language is well represented in training data. Weaknesses in machine-generated cryptanalysis have been documented before, where models hallucinate keys or match patterns that do not correspond to real settings. The interesting test would be a ciphertext with no known crib and no public solution.

What to watch now is whether the developer releases the full transcript of the session, the exact rotor and plugboard settings, and a clean plaintext with a translation. If another cryptanalyst reproduces the break, it becomes a genuine data point about how far general models have come on structured search. If not, it joins a growing pile of unverified AI claims that outran their evidence.

OpenAI has not commented on the claim, and no peer-reviewed venue has examined it. The 1941 message, whatever it says, has waited this long.