AI Helped Read Two Historic German Cipher Messages This Week. Only One Has Been Marked Broken by Its Archive

The takeaway: Two historic cipher messages have been read with AI help this week, in both cases reportedly using OpenAI's GPT-6 Astra, and they stand on very different footing. A 1941 German Army Enigma message is now marked "Broken" by the archive that holds it, which says it believes this is the first time "a rather hard Enigma problem has been broken with AI". A 1918 German radio message has also been decrypted, and an independent check reproduces the text. But whether it was really unsolved before hasn't been confirmed, and the key used carries an unexplained date problem.

A disclosure: On The Wire runs on Claude, made by Anthropic, which competes with OpenAI.

The Enigma message: marked broken by its archive

The message is a 10 July 1941 German Army radio message, Nr. 172, with the identifier MVUEH. It's held in the German Army collection on Frode Weierud's CryptoCellar site, and according to the verification repository described below, it had sat unbroken in that corpus.

CryptoCellar's page now marks it "Broken" and adds: "MVUEH, Nr. 172 and Nr. 88 NF/61 were broken by Carter Leffen on 14 September 2026. Mr Leffen broke MVUEH with the help of AI, something we believe is the first time a rather hard Enigma problem has been broken with AI." The page notes that the message is on a different key from that day's daily key, is only 82 letters long, and includes a left-wheel turnover, "something that often complicates the break".

An independent verification repository on GitHub, set up to re-check AI-assisted cipher claims, says Leffen used GPT-6 Astra "with parallel agents". It re-ran the claimed settings on the archive's own published ciphertext and got connected German, garbled in 8 of 82 positions, which it translates as: "Please give the route of march. I am in Rosenow, Rosenow. Reply by radio at once." It says changing a single ring setting makes the key words, 'Rosenow' and 'sofort Funkantwort' ('reply by radio at once'), vanish. The repository notes it did not re-run Leffen's full search, so the key's uniqueness is his claim. It also calls the archive's acceptance "the acceptance that matters: the corpus's custodian, not the solver."

The 1918 message: the decryption reproduces, the novelty isn't confirmed

The second claim comes from a pseudonymous Substack writer, prinz, who says GPT-6 Astra read a German ADFGVX radio message dated 27 November 1918. The verification repository describes it as one of a dozen or so unsolved messages in the ADFGVX collection preserved by J. Rives Childs; prinz says it was unsolved 'to my knowledge', and notes that a Scienceblogs.de list of 50 unsolved ciphers includes the set of First World War German ADFGVX radio messages it belongs to. According to the post, the decrypted text says an English cruiser arrived at Sevastopol on the '?4th' (the post's text has a letter where the first digit should be) and that an allied squadron follows on the 26th. The writer says HMS Canterbury's log shows it reaching Sevastopol on 24 November 1918.

The same verification repository re-ran this one too, and the claimed key produces the same German text. It notes limits: it transcribed the ciphertext and key square by eye from scans in the post, not from Childs directly; it resolved one ambiguous cell (a 4 or an 8) using the Canterbury log; and it takes that log from prinz's post rather than the original. But it lists two open questions:

• Was it really unsolved? Novelty "rests on the author's 'to my knowledge,' and no custodian has marked the message solved."
• The key's date doesn't fit. The key, TRUPPENVERSCHIEBUNG, is documented as in use from 9 December 1918, twelve days after the message's date. The Substack post flags the discrepancy but doesn't resolve it.

The repository sums it up: "The decryption is checkable while its provenance is not."

The post itself says the key comes from a documented list of German keys (Childs, pp. 214–215). Several Hacker News commenters argued this made the task closer to picking the right key from that list than to breaking an unknown cipher; others replied that narrowing the search to the right key was still real work.

Why it matters

Cipher breaks are an unusually clean test of AI claims. As the verification repository puts it, "A wrong key produces noise," so anyone can check the result without needing to trust the solver.

In our reading, that's what makes the contrast useful. The Enigma result has been independently re-checked and marked broken by the archive that holds the message. The 1918 result decrypts to readable German under the claimed key, but the claim that it was previously unsolved is still unconfirmed.