GPT-6 Astra breaks 1941 Enigma message MVUEH that stayed unbroken since 2005
- Carter Leffer asked GPT-6 Astra to try the unbroken messages on the Crypto Cellar Research page, and the model recovered the key and plaintext for German Army message MVUEH (Nr. 172, 10 July 1941). Frode Weierud validated the break on 15 September 2026 after it had resisted solution since 2005.
- The MVUEH key uses wheel order 253, while the daily key and the 2017 SIPVX key for 10 July 1941 both use 512, and the plug connections and ring setting match neither of them.
- The recovered plaintext is nearly identical to message Nr. 173 SIPVX, broken by Alex Shovkoplyas in 2017. MVUEH runs 82 letters against SIPVX's 94 because Bitte was mis-enciphered as Btte and the signature Waschbusch appears twice in SIPVX.
- Astra built its own Python and C++ Enigma simulator and Bombe, then used the repeated place name ROSENOW ROSENOW as a crib to search out the correct key.
- Two obstacles may explain the long resistance: transcription errors in the copied MVUEH ciphertext, and a left-hand wheel turnover at the 72nd letter. Weierud says the model's logs also cite Bundesarchiv files RS 3-3/20a and RS 3-3/63b that never appear on the site, so how it found them is unknown, and its two days of work equals weeks or months of his own.
Hacker News opinions
How many of these news articles are we going to get? This needed no skill and no imagination, it looks like dumb luck. Now an army of know-nothings points models at old forgotten tasks to get 15 minutes of attention.
Half of all useful human inventions came out of dumb luck too. That does not make them uninteresting to me.
This wasn't possible just months ago. For anyone not following closely it is still newsworthy, though probably not for long.
AI is great at pattern matching and finding solutions by assonance. I agree these are not thrilling discoveries, and I dislike them mainly because they will be used to push the next IPOs, Anthropic and OpenAI.
If human skill and imagination are your bar for interesting, you are going to be permanently bored fairly soon.
First time I'm seeing this. I had heard about the WWI German radio cipher crack, are you confusing the two?
I'd be curious whether these models can create new forms of unbreakable encryption themselves.
They seem to be better at solving concrete problems than designing new things, at least for now.
They are trying to generate hype before the IPO so they can cash in before the bubble bursts.
I think it was Roon who said it months ago: too many people are not emotionally prepared for if it's not a bubble.
Maybe I'm wrong, but this looks like a strange form of competitiveness. People don't care about the advancements themselves, only that the advance is some kind of cheating that shouldn't count. Humanity is unsettled by AI and responds with denial.
I wouldn't be surprised if humans cracked it and then OpenAI bought the solution to claim it was GPT's doing.
The actual plaintext is BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH, roughly: Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio. Waschbusch.
The Astra part about the crib feels extremely underexplained. Why would it use ROSENOW ROSENOW as a crib, was repeating the place name common in these messages? A single ROSENOW would work just as well as a crib.
Do we know why those specific messages used different keys? Almost identical content with the key changing first partly and then everything looks more like an error or a test than special high-security keys.
I want an AI trained only on knowledge through 1903, then see if it can come up with E=mc^2.
Even the Millennium Prize Problems have become benchmarks for model companies. Intelligence is now a product you quantify and buy with electricity, and that feels awful.
You can't really replace the smartest individuals, because a person who is not smart can't do much with the AI. What is happening is that access to expertise is becoming commodified.
Solving obscure puzzle samples that maybe zero humans ever attempted, mostly by pattern matching known solutions to similar puzzles, is not intelligence. DeepBlue beat the best humans at a puzzle-like task last century. Meanwhile I deal every day with reams of horrific code that technically compiles and has mistakes any human would be fired for.
Well, let them have the prize problems. They generate media hype and then the companies move on, because "AI solved in 3 days what humans couldn't in 100 years" won't repeat for problems the AI has to invent first.