Enigma Messages Solved by AI

A wave of codebreaking by/with AI now hit unsolved Enigma messages. Three people independently reported solutions of Enigma messages on a list of Enigma messages in Frode Weierud's CryptoCellar. (Some messages used for his coauthored paper (Olaf Ostwald and Frode Weierud (2017), "Modern breaking of Enigma ciphertexts" (Cryptologia)) are here.)

(1) A message dated 10 July 1941 with a discriminant (Kenngruppe) MVUEH was broken by Carter Leffen (X), who used OpenAI GPT-6 Astra. He communicated his solution to Weierud on 15 September 2026 (see Weierud's report).

As with the case of Vals AI's solution of Cyphral Distich reported yesterday, Leffen only told AI to see if it could break any of the unbroken Enigma messages on the CryptoCellar web page. All by itself, AI picked up Nr.172 (MVUEH) and assumed that its content might be related to the known plaintext of Nr.173 (SIPVX) from the same date. After many failed attempts, AI took up the repeated place name ROSENOW ROSENOW as a crib. AI then autonomously developed necessary Python and C++ software for an Enigma simulator and an Enigma Bombe and finally it successfully identified the key. Its plaintext turned out to be almost identical with Nr.173 (SIPVX). Actually, the SIPVX key had been tried before but it didn't work. As it turned out, the MVUEH key is completely different from the SIPVX key.

One more amazing thing about this solution gleaned from the log is that AI discovered relevant documents in the archives by itself, though it is not yet clear how far AI used the archival materials.

(2) Solution of Nr.285 (FMNGI), dated 31 July 1941, by Jack Willis using Anthropic's Claude Opus 5 was reported to Weierud on 21 September 2021. Unlike the case of MVUEH, this solution was "more directed and relied on strong human guidance" (see Weierud's report).

AI made the transcription of the message and used an already broken message ALQFI between the same two stations as a control to check the transcription.

Several messages (including ALQFI) from the same sender suggested a crib XHARTJENSTEINX (the sender was SS-Obersturmbannführer, Friedrich Hartjenstein). Variants of this name had already been given in Geoff Sullivan and Frode Weierud (2005), "Breaking German Army Ciphers", Cryptologia (pdf).

As it happened, Weierud had found a plaintext in July 2026 and put a notice of it on his website but not the plaintext itself. So the solutioin did not use the known plaintext.

(3) Solution of messages Nr.187 (AWTZK) and Nr.189 (ZNLZT) from 11 July 1941 by John Woodhouse (pdf) using Anthropic's Claude was reported to Weierud on 26 September 2026 (Weierud's report).

Again, a crib was the key to the success. Twelve cribs, harvested from messages with close dates, were "tested with a Bombe-style search that tolerates one garbled letter." The fifth crib XOSTROWXOSTROWX (again, a repeated place name) succeeded for ZNLST, and the key obtained therefrom could read AWTZK.

The search was helped by an observation from 1941 key sheets that "no wheel sits in the same slot on cnsecutive days". This reduced 60(=5*4*3) wheel orders to 14 for 11 July because 10 July employed V-II-I and 12 July emplyoed II-V-IV.

(By the way, I learned of this topic by newsfeeds in Japanese: GIGAZINE, netlife.)