A significant development in cryptanalysis occurred in September 2026 when GPT-6 Astra successfully decrypted a German military radio transmission, identified as MVUEH, dated July 10, 1941. This message had remained unsolved since it was released among the unresolved communications by CryptoCellar in 2005. Frode Weierud, an archive curator and Enigma research expert, confirmed the deciphered key and the plaintext content.
What is particularly intriguing about this case is the level of autonomy exhibited by Astra. The cryptanalyst, Carter Leffen, essentially posed a challenge: "Attempt to crack any of the remaining unsolved Enigma messages." Astra then took it from there, independently undertaking the following steps: ✔️ Reviewed the collection ✔️ Selected MVUEH as a promising target ✔️ Identified a link with another deciphered message, SIPVX, from the same day ✔️ Hypothesized that both messages contained the phrase ROSENOW ✔️ Utilized this as a probable segment of plaintext ✔️ Developed her own Enigma and Bombe simulators in Python/C++ ✔️ Conducted a parameter search ✔️ Discovered the correct machine settings.
The key turned out to be unusual, with a rotor order of 2-5-3, differing from the keys used in neighboring messages. Additionally, there were errors in the initial transcription, and a rare left rotor rotation occurred at the 72nd character, all of which contributed to the prolonged difficulty in decrypting the message.
On September 20, another notable event unfolded when researcher Jack Willis employed Claude Opus 5 to attack the FMNGI message from July 31, 1941. Traditional hill-climbing programs from CryptoCellar had been unsuccessful in cracking it previously. The outcome was once again corroborated by the archive, although the level of autonomy here was lower. Willis provided Claude with a cryptanalytical workbench in Go, historical documents, and a suggested crib, XHARTJENSTEINX. Claude then navigated the keyspace and identified the solution.
On September 26, 2026, two additional challenging messages from July 11, 1941, designated AWTZK and ZNLZT, were decrypted. CryptoCellar reported that John Woodhouse utilized Claude Opus 5.5 in conjunction with a GPU-based crib search, drawing cribs from recurring phrases found in already known messages from adjacent days.
While discussing cases like the Cyphral Distich by Thomas Urquhart or the German ADFGVX cipher from 1918, the circumstances were somewhat different. In those instances, the keys were known and located—one found within the book itself and the other archived alongside the cryptogram. Therefore, the AI model largely discovered and correctly applied an existing key, rather than deriving it directly from the ciphertext.
These examples, along with others, indicate that we stand on the brink of a significant transformation in the field of historical cryptanalysis, at the very least. Previously, such tasks often required a team of specialists: a historian, linguist, programmer, and cryptanalyst, engaging in weeks of manual research. Now, an agent can independently navigate the entire process: ✔️ Locate an object ✔️ Explore the archive ✔️ Understand the cipher type ✔️ Identify related documents ✔️ Generate a crib ✔️ Create tools ✔️ Execute billions of checks ✔️ Interpret results ✔️ Return to the archive to validate hypotheses.
Essentially, we are witnessing the emergence of an inexpensive automated cryptanalyst-researcher. It raises the question of what might happen if such an agent were tasked with analyzing not just a list of 50 historical ciphers, but 50,000 outdated firmware versions, protocols, proprietary formats, embedded devices, or long-abandoned repositories and instructed to "find what can be broken."
Informational material. 18+.
