On 15 September 2026, Carter Leffer contacted cryptanalyst Frode Weierud to validate a break of the German Army Enigma message MVUEH, dated 10 July 1941. The message, sent by a station with tactical callsign 2ny and received by the SS-Totenkopf Quartiermeister radio station, had resisted all decryption attempts since 2005. The break was not performed by a human cryptanalyst. OpenAI's GPT-6 Astra did the entire thing on its own. The story of how is what matters. Leffer pointed GPT-6 Astra at the Crypto Cellar Research web page listing unbroken Enigma messages and told it to see what it could do. The model triaged the candidates, selected MVUEH as the most promising target, and noticed that the plaintext of a related broken message (Nr. 173, SIPVX, broken in 2017 by Alex Shovkoplyas) might provide a crib — a known-plaintext anchor for attacking the cipher. It then settled on the repeated place name ROSENOW ROSENOW as its attack vector. What followed was not pattern-matching or lookup. GPT-6 Astra developed its own Enigma simulator and Enigma Bombe software in Python and C++, then conducted a systematic cryptanalytic attack. The recovered key turned out to be completely different from the daily key used for all other messages on 10 July 1941 — wheel order 253 instead of 512, different plug connections, different ring settings. The plaintext was nearly identical to that of message Nr. 173, differing by only twelve letters due to an encipherment error (Btte instead of Bitte) and a repeated sender signature (Waschbusch). The break also resolved longstanding mysteries. The MVUEH ciphertext transcription contained several errors, and the Enigma's left-hand wheel turns over at the 72nd letter — a rare event known to complicate cryptanalysis. These factors likely explain the message's 21-year resistance. Perhaps the most striking detail is the archival research. GPT-6 Astra discovered a note on the Crypto Cellar website about radio message collections at the German Bundesarchiv. Logs show it identified the correct file references — RS 3-3/20a and RS 3-3/63b — which are not published on the website. Weierud confirms these references are accurate and notes he personally spent several weeks researching them. The AI accomplished in two days what a seasoned researcher measured in weeks to months. The implications extend well beyond historical cryptography. This is an autonomous research agent performing multi-step reasoning across domains: triaging targets, forming hypotheses about plaintext relationships, writing bespoke software tools, conducting the attack, and pursuing independent archival leads. It did not merely execute a known algorithm. It chose an approach, built the tools for that approach, and adapted when the problem's peculiarities — transcription errors, unusual key configurations, rare wheel turnovers — would have derailed simpler methods. Weierud, a veteran cryptanalyst, calls the result "simply amazing." The analysis of GPT-6 Astra's logs is ongoing, and more details about its reasoning process are expected. What is already clear is that the system operated at a level of professional competence that maps directly onto the workflow of a skilled human analyst — target selection, crib identification, tool development, systematic attack, and archival verification — compressed from weeks into roughly 48 hours.
