ADFGVX was a German field cipher from the First World War. It turns each letter into a two-letter pair built from a six-letter alphabet, then reorders the columns of the result using a key word — substitution followed by transposition. Hundreds of German radio messages have been broken over the decades, and about a dozen remain unsolved on a public list. prinz ran GPT-6 Astra at one of them, transmitted on 27 November 1918, and reports a readable German plaintext.

What the post reports:

  • The recovered message, in the model’s reading: an English cruiser arrived at Sevastopol, and an allied squadron follows on the 26th.
  • The key was not invented. The model used “TRUPPENVERSCHIEBUNG,” an existing German key word listed in published sources and described on pages 214–215 of J. Rives Childs’s history of German military ciphers.
  • The anomaly the post flags: that key was documented as entering service on 9 December 1918, two weeks after this message was sent. Astra’s hypothesis is that the operator used it early; the post says the discrepancy is unexplained.
  • The self-check: the model then looked for corroboration in naval records and found HMS Canterbury’s original logs, which place a cruiser at Sevastopol on 24 November 1918, with an allied squadron arriving on the 26th.
  • The author’s own framing is modest — “a minor (but I think really cool) result and illustration of the capabilities of this model,” noting he is unaware of this particular message having been decoded before.

So the substance is a model applying a documented decryption procedure to a documented key that human cryptanalysts had reason not to try, and getting a result that matches the historical record. That is a real use of the tool, and it is also the kind of claim that needs someone else to check it.

The 134-comment thread on Hacker News is that checking, informally.

What the thread adds

  • meindnoch — the correction at the root of the headline: “So it used a known key. It didn’t come up with a key from thin air. The only gotcha is that apparently this key was used two weeks earlier than it was documented (maybe the operator was using the wrong page from the codebook?).”
  • grey-area — makes the same point as a complaint about framing: “Using an existing published key, which people hadn’t tried because the message was sent before the key was supposed to be used. This headline is misleading.”
  • eis — went to the source and found an awkward gap. They asked Astra to describe the cited pages 214–215 of the Childs book: “It said it can’t and it can’t find this book online either.” flats replies that the book appears to be unpublished and held by a museum — which leaves the citation, not the ciphertext, as the weak link.
  • 93po — an adjacent story about how models handle disputed sources. ChatGPT told them a previous cipher solution was “disputed online” and made the dispute sound convincing; the dispute turned out to be an AI agent’s own transcription error, counting book passages that continued onto the next page as too short. Their conclusion: “Annoying that ChatGPT can cite sources like this without being able to properly weigh their reliability.”
  • iltk, answered by binlog — the forgery hypothesis, stated and then mostly dismissed on its own terms: could the model have reverse-engineered a fake key against public ship logs? binlog: forging a key that produces a message matching the ship’s arrival day “would be significantly more impressive than just cracking it.”
  • durdn — the state of play on LLM cryptanalysis, written as a litany of moving goalposts, from “LLMs can’t do cryptanalysis… they can barely solve toy substitution ciphers” through acknowledged results on reduced-round AES and LEA, ending on “Wake me up when it breaks full AES.”
  • aaymeloglu — the generic version: after a previous HN cipher post they pointed Astra and Fable at other unsolved messages and found “plenty of low hanging fruit,” which suggests the value here is breadth of patient search rather than a new technique.
  • jgrahamc, correcting FabHK — ADFGVX is not a simple substitution cipher; the substitution is only the first step, and “the second step is then a keyed columnar transposition.” zamadatix supplies the general answer for why no human had done it: not every value of AI has to be superhuman, sometimes it is just doing labor nobody thought was worth the effort.

The question the thread kept asking

Whether the answer is genuinely new. The post says the author is unaware of a prior decode of this message, and the thread splits on what that means: sehw (“Or, it found a human that solved it in the dataset and stole the solution”) and nojvek (“GPT-9 solves X. ‘Oh GPT-9 found a solved solution on internet and claimed as its own.’”) assume dataset provenance, iltk allows for fabrication, and pmarreck pushes back on the cynicism — “there’s plenty of evidence that they can do original work.” Nobody in the thread demonstrates either way, and the article’s own evidence for novelty is the author’s awareness plus a ship-log date match. That is thinner than the headline, but it is not nothing: the log check is the part of the claim a reader can actually repeat.

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering here is HN’s own ranking rather than a vote, and this is a slice of the thread — not a consensus.