OpenAI's GPT-6 Astra decoded an Enigma message from 1941 that human cryptanalysts had left unsolved since 2005

An OpenAI model chose its own target from a public archive of unsolved wartime ciphers, wrote the software it needed to attack the target, and returned a plaintext in about two days. Frode Weierud, who runs Crypto Cellar Research, wrote up the run on 19 September 2026 and calls the result "simply amazing" in the same post.

The target was a German Army Enigma message from July 1941 known as MVUEH, sent by the SS-Totenkopf Division during Operation Barbarossa and unsolved on Weierud's own site since 2005. Weierud gave GPT-6 Astra access to the archive and asked it to find a message worth attacking. Astra picked MVUEH because it saw a possible connection to a related solved message called SIPVX and thought a common German word from that message could open the key.

Why MVUEH had stayed unsolved for two decades

Two features of MVUEH had defeated the standard hand and computer attacks used by hobbyist cryptanalysts. The operator made typing mistakes in the ciphertext, and the left rotor turned over in a rare position at character 72, so a search that assumed a normal rotor pattern missed the key. Weierud writes that the ciphertext was already in the archive at the level of quality he expected of a hard case.

The model wrote its own Bombe

Rather than call an existing library, Astra built its own tools for the attack. From the writeup: "After developing the necessary Python and C++ software for an Enigma simulator and an Enigma Bombe, GPT-6 Astra started a thorough break with the ROSENOW crib." It then produced the plaintext: "Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio. Waschbusch." Waschbusch signs off as the division radio operator, which matches the archive's records.

The choice of clue is worth noting on its own. In Enigma work, a crib is a guess at part of the plaintext, and the guess here was the town name ROSENOW, drawn from the sister message SIPVX. A guess like that comes from reading around the archive, and no packaged test set would have handed it to the model.

Weierud's verdict, in his own words

Weierud's account of what the model did is a first-person report from the person who runs the archive. The direct quote in the writeup: "GPT-6 Astra is behaving like a very professional cryptanalyst and archive researcher. What it achieved in two days would take a human researcher weeks or even months." He does not claim the model has replaced him, and the writeup describes the interaction as a research assistant that he could point at the archive.

For a team building with agent frameworks, the run is one worked example of a pattern often shown on made-up problems and rarely on a real archive. The model picked the target, built the tools it needed, tied a specific historical clue to a specific ciphertext, and produced a plaintext that Weierud could check. A team pointing an agent at its own store of hard, unsolved cases, log files, error reports, or half-decoded file formats, can read this as the current best case for what a well-directed model can finish on its own. The story reached 687 points on Hacker News, which is where a working developer audience saw it.

Source

Frode Weierud, Crypto Cellar Research, The MVUEH Break, updated 19 September 2026. Discussion: Hacker News thread 49801324.