OpenAI introduced its new AI model Astra in early September. Since the debut, observers have been testing the model's capabilities. Researchers at the independent laboratory Crypto Cellar now report that GTP-Astra decrypted an 85-year-old Enigma ciphertext, the so-called MVUEH message, in two days.
What was the message and when was it sent?
According to Crypto Cellar, the MVUEH message was sent by the German High Command to the SS-Totenkopf division on 10 July 1941, during the Second World War. The ciphertext remained unsolved for decades; it was posted online in 2005 but, the lab says, its true plaintext stayed unknown until now.
How did Astra reach the decryption?
Crypto Cellar's report states that Astra performed the task autonomously after researchers asked whether it could attempt to break one of the previously unsolved Enigma messages published on the Crypto Cellar website. The system selected message number 172 because it judged the message might be related to message number 173, marked SIPVX.
Expecting the repeated place name “ROSENOW ROSENOW” as a likely plaintext clue, the model proceeded to develop software in Python and C++ for an Enigma simulator and a code-breaking machine. According to the lab, Astra completed the decryption within two days; the researchers estimate the same work would have taken human analysts weeks or months. Crypto Cellar is currently investigating precisely how the model carried out the task and achieved such rapid results.
Ciphertext and plaintext
The initial encrypted text published by the lab was:
ICRVSORMCCWQTATYEVFXDBZGGSNXWLPSYWZYTCBSWULRTBZCVGODVJUSLSOOMJQJZSXSEBZPEYMDNXJYTC.
The solution reported by GTP-Astra was:
BTTEUMANGABEDESMARSQWEGESXBEFINDEMIQINXROSENOWROSENOWXSOFORTFUNKANTWORTXWASCHBBSCH.
Crypto Cellar provides an approximate English translation of the plaintext: “Please provide the route. I am in Rosenow, Rosenow. Request immediate radio reply.” The lab also notes that a spelling error at the very beginning of the original text complicated the codebreaking.
Why this matters
If Crypto Cellar's claim and the presented solution are accurate, the result would illustrate that modern large language models and AI systems can be applied to historical cryptanalysis and can solve complex cryptographic problems in short timeframes. The laboratory emphasizes that it is continuing to examine the methods and the reliability of the outcome.
Crypto Cellar has published a technical report on the decryption process and further details on its website.
