Computer pioneer Alan Turing is best known for his experiment of the same name, which tested whether artificial intelligence could be distinguished from human intelligence. But the more important test he faced in his life was cracking the Enigma code used by Nazi Germany during World War II.
Now, two cryptanalysts have announced that they have used an LLM built by OpenAI and Anthropic to decipher two long-unsolved messages encoded on the Enigma machine.
Turing and his code-breaking team built an early computer called Bombe that allowed Britain to translate Enigma messages during the war, but a small number of archived messages remain uncorrupted, usually due to incorrect transcription or errors by the original encoder.
Developer Carter Leffen simply told OpenAI’s latest model, Astra, to search a database of Enigma messages for uncorrupted messages and decrypt them. The model was able to do just that after conducting its own archival research, finding context clues, building a simulator of the Enigma machine, and ultimately recovering the plaintext of a message that has baffled researchers since 2005. He was even able to use Astra to build an interactive website that explains the entire problem.
Frode Weierud is a former electrical engineer with a long-standing interest in cryptography and maintains the website Crypto Cellar, which includes a database of various resources, records, and messages. Weield tested Leffen’s solution last week and said he was “in awe.”
Especially given the long time the AI agent takes to answer the question at hand, the Astra model’s logs include discussions about messages archived in “private collections” not hosted by Weierud. He doesn’t yet know if the model had access to them, but speculates that they may have been shared somewhere online by another researcher, or that the model had access to German government archives.
“GPT-6 Astra is behaving like a highly specialized codebreaker and archival researcher,” he wrote. “What we accomplished in two days would take a human researcher weeks or even months. Personally, I spent several weeks examining the Bundesarchiv files referenced by GPT-6 Astra.”
On September 21, another codebreaker and cybersecurity executive, Jack Willis, contacted Weierd and said he had used Anthropic’s Claude Opus 5 model to crack another unresolved message. Willis provided Claude with more guidance, and Claude was eventually able to decipher the message using the known signature of a particular officer’s name.
Weierud says there are only seven Enigma messages left that have not been deciphered, and one message whose plaintext is known but whose code has not been deciphered.
It probably won’t last long.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
