A Claimed AI Break of an Enigma Message Still Needs a Re-encryption Test
A high-scoring Hacker News post says an AI model has read an Enigma message unsolved since 2005. Cryptanalysis is one of the few fields where a claim can be checked by anyone with a simulator, so the next step is narrow and public.
The claim, as far as it goes
A link to cryptocellar.org titled “OpenAI GPT-6 Astra breaks Enigma message that has resisted solution since 2005” climbed Hacker News’ Best list, drawing 727 points and 439 comments. The page it points to is named “the-mvueh-break.”
That is nearly the whole of what the source supports: someone has published a claim that an Enigma-encrypted message, unsolved since 2005, has now been read. The model named in the headline, the conditions it ran under, the method, the plaintext itself — none of these have independent backing in the material at hand. The claim may turn out to be correct. For the moment it is a claim, and cryptanalysis happens to be a field where claims are unusually cheap to check.
The test is arithmetic, not rhetoric
Enigma is a reciprocal machine. Run the plaintext back through the same rotor positions and plugboard settings and the ciphertext reappears; one operation serves as both encryption and decryption. So a claimed break arrives with its own verification procedure attached. Publish the settings and the plaintext, and anyone with a simulator can re-run the message and compare. If the letters line up, the break stands. If they don’t, nothing else matters.
There is a cheaper filter too. Because of the reflector, Enigma never enciphers a letter as itself. If any position in a proposed plaintext matches the ciphertext letter at the same position, the plaintext is wrong — no machine, no key, no argument required. That constraint is worth more than a passage that reads like German.
The hard part is ambiguity. Two decades of hobbyist attacks have swept much of the plausible key space, and among what remains there will be settings that produce fragments of passable German. Deciding which candidate is real depends on the coherence of the whole message, on period phrasing, and on whether the settings reproduce the ciphertext exactly. Language models are strong at the first of those. They generate plausible German fluently, which is also the risk: fluency is not correctness, and a plaintext completed by a model, with no re-encryption check behind it, can be an elegant guess.
A crowded comment section is not evidence
439 comments say the thread is interesting, nothing more. Unsolved Enigma traffic has served as a public puzzle set for amateur cryptanalysts for years, and anyone who announces a solution invites the same questions: what are the settings, where did the plaintext come from, can it be reproduced. Those questions are not gatekeeping. They are the only acceptance procedure this kind of work has.
Models in cryptanalysis are nothing new; pointing one at a corner of the search space and asking it to rank candidates is a natural use. Guessing fast and guessing right are different things. A message that has absorbed repeated attacks since 2005 is telling you something — either it needs an external clue nobody has found, or the candidate space is more crowded than it looks. A model can shorten the search. It cannot decide, on its own, which plaintext is the message.
What happens next
If the cryptocellar page publishes complete key settings and plaintext, verification will be quick and public, and the argument ends with a re-run. If it doesn’t, the discussion stays where it is, regardless of which model name is in the headline.