OpenAI's GPT-6 Astra has recovered the settings behind an 82-letter German Army Enigma message sent on July 10, 1941. The message, known by the indicator MVUEH, had sat unbroken on a public research archive for 21 years. The interesting part for developers is not the history. It is that the model wrote and ran its own cryptanalysis software to get there.
Enigma was the cipher machine the German military used in the Second World War. Its settings changed daily, and breaking a message means recovering those settings.
What the model actually did
The CryptoCellar account of the break describes the sequence. GPT-6 Astra read through the unbroken messages held on the site and picked MVUEH as the most promising target. It then noticed that the plaintext looked connected to a different message, Nr. 173, indicator SIPVX, which researchers had already solved.
From there the model wrote software. It built an Enigma simulator in Python and in C++, plus an implementation of the Bombe. The Bombe was the electromechanical machine built at Bletchley Park to search Enigma settings. The model then ran that search.
Bruce Schneier's summary describes the result as "a thorough break with the ROSENOW crib". A crib is a guess at a piece of the original text. If you know some words the message almost certainly contains, you can test settings far faster.
Where the crib came from
That crib is the reason to read the autonomy claim carefully. The repeated place name ROSENOW ROSENOW came from SIPVX, the related message that human researchers broke in 2017. Schneier's write-up says the model "did it entirely on its own", and in the same breath names its reliance on that crib.
Both things are true. The model chose the target, spotted the link between the two messages, wrote the tooling, and ran the search without being told how. It did so standing on an existing human break of a sibling message.
Schneier's post adds one more caveat. The team is "still analysing the GPT-6 Astra logs to see exactly how it executed the break". The full method is not yet public.
The verification
Frode Weierud, the cryptanalyst who runs the CryptoCellar archive, checked the result. CryptoCellar's account says he confirmed that the model "had found the correct key and plaintext", and calls the achievement "simply amazing".
The recovered key was not the expected one. CryptoCellar reports a wheel order of 253, where 512 was the order otherwise recorded for that day. The wheel order is which rotors sit in which slots inside the machine. The recovered plaintext is nearly identical to SIPVX, with differences the account puts down to encipherment errors made at the time.
CryptoCellar also explains why the message resisted for so long. The ciphertext as transcribed contains errors, and the message involves a rare turnover of the left-hand wheel. Both break the assumptions a straightforward search depends on.
What this means for developers
The transferable result here is a workflow, not a cipher. The model surveyed a problem set, picked the tractable instance, wrote a correct simulator in two languages, wrote a search program against it, and ran it. That is ordinary engineering work, done end to end, on a problem with a verifiable answer.
Note which conditions made it work. The answer was checkable, a related solved case existed to supply a crib, and the search space was bounded by physical machine constraints. Where your own problem has those three properties, this pattern is worth trying. Where it does not, a model cannot tell you it has failed.
The verification step is the part not to copy loosely. A named human expert confirmed the key before anyone published the claim. That is how this result stayed a result, rather than joining the pile of plausible-looking machine output nobody checked. CryptoCellar's account says the two days of work would take a human researcher weeks or months, which raises the value of fast expert review rather than lowering it.
The security reading is narrower than some coverage suggests. This was a 1941 cipher with a tiny keyspace by modern standards. It says nothing about the strength of anything you are shipping. It does follow OpenAI rating Astra its first Critical cyber model five days earlier, which is the more concrete signal for anyone tracking this capability.
This article was first published on Tech AI Wire.
Also available in
Deutsch · 日本語 · Français · Español · Português
Related on Tech AI Wire
Sources
- The MVUEH break - CryptoCellar
- GPT-6 Astra Breaks an Old Enigma Message - Schneier on Security
Top comments (0)