Artificial intelligence systems from OpenAI and Anthropic were able to decrypt messages encoded by Enigma machines, which had remained unsolved for many years. The models used in the work were GPT-6 Astra and Claude Opus 5.
These results attract attention because Enigma was used by Nazi Germany during World War II and was linked to the efforts of Alan Turing and other cryptanalysts to break its codes. According to Frode Weierud, seven messages remain undeciphered.
Developer Carter Leffen asked Astra to find another undeciphered Enigma message in the database. The model analyzed the files, studied historical clues, and modeled the machine to attempt to reconstruct the original text. This message had been unsolved since 2005, and after two days of work, Astra provided an answer. Leffen also used this model to create an interactive website about this issue.
More Details
The solution was analyzed by Frode Weierud, a retired electrical engineer and cryptography researcher who runs the Crypto Cellar website. He confirmed the result and expressed admiration for the system's work. Astra's records also revealed an open question: the model mentioned messages from a 'private collection' that are not in Weierud's database. The researcher does not know if the AI accessed this material, speculating that the documents might have been passed on to another researcher online or reside in public German government archives.
Weierud noted: "GPT-6 Astra behaves like a very professional cryptanalyst and archive researcher." In his opinion, what the model achieved in two days would require weeks or months of human labor.
On September 21, cryptanalyst Jack Willis contacted Weierud to report that he used Claude Opus 5 to attempt to decrypt another Enigma message. The process was different: Willis provided the model with more specific instructions, including a clue based on a known officer's name. Thanks to this information, Claude was able to obtain the message text.
Both cases demonstrated different strategies: Weierud claims there are seven undeciphered Enigma messages, as well as an eighth whose original text is known, but the code remains unbroken. These two findings reduce the number of historical messages still resisting decryption. The application of new AI resources to such research may mean that the remaining records will not remain unanswered for long.